betaFeedback

WorksComputer Science

LiDSeg: Linear DehazeFormer Guidance for Efficient Semantic Segmentation in Adverse Weather Conditions

Abstract

Robust semantic segmentation in adverse weather is essential for safety-critical autonomous systems. Existing restore-then-segment pipelines introduce latency and prioritize pixel-level restoration over semantic discriminability, while domain adaptation methods often face representation bottlenecks and computational complexity. We introduce LiDSeg, a framework that uses reconstruction-informed guidance to improve segmentation under adverse weather. During training, a lightweight guidance encoder is jointly optimized with the segmentation branch using segmentation and auxiliary reconstruction objectives. At inference, the reconstruction decoder is discarded, while the guidance encoder is retained to extract weather-robust features. A lightweight gated fusion module integrates these features into the segmentation branch, preserving clear-weather accuracy while improving adverse-weather robustness. On Foggy Cityscapes, LiDSeg limits the clear-to-dense-fog mIoU degradation to 2.5-2.8 percentage points, reducing the baseline performance drop by more than fivefold. On the official ACDC test server, LiDSeg-PIDNet achieves 75.07% pooled all-condition mIoU, comparable to supervised HRNet-W48 at 75.0%, while remaining substantially smaller. The gains generalize zero-shot to Foggy Driving and Foggy Zurich without training on real-fog images. LiDSeg adds only 1.9-3.1M parameters across several models, with total model sizes of 5.2-39.5M parameters and no inference-time reconstruction overhead. It delivers real-time inference on the NVIDIA Jetson AGX Orin. Code: https://github.com/yobsanb/LiDSeg

Introduction

Semantic segmentation in adverse weather conditions is a fundamental challenge for diverse autonomous systems and safety-critical applications [36]. Despite significant advances in clear-weather datasets, such as Cityscapes [8], model performance declines in adverse weather conditions due to weather-induced artifacts that corrupt the semantic features essential for scene understanding [2, 54, 59]. Common methods for addressing this problem include sequential restore-then-segment pipelines that explicitly restore degraded images before segmentation [33, 49]. While improving the visual quality of the restored image, these methods introduce additional inference latency and emphasize pixel-level restoration objectives, which do not necessarily enhance semantic discriminability [44]. Another method is leveraging domain-adaptation techniques to align feature distributions across weather domains [11, 19, 24, 29, 36, 40, 45]. However, this method often leads to representation bottlenecks and limits the robustness gains [29]. To tackle these challenges, LiDSeg employs a modular framework with asymmetric training and inference that integrates weather-robust guidance into a segmentation network (Figure 1). During training, the auxiliary reconstruction loss,

Figure 1. LiDSeg architecture. The degraded image I is processed by a guidance branch (Linear DehazeFormer encoder with training-only decoder) and a segmentation branch.

computed on paired adverse and clear images, jointly optimizes the shared guidance encoder together with the segmentation loss, so the encoder learns features robust to the distribution of weather corruptions. During inference, the reconstruction decoder is removed for efficiency, whereas the LGF module selectively fuses the encoder features into the segmentation network to enhance robustness in adverse weather conditions without compromising clear-weather accuracy. This design augments a standard segmentation network with minimal additional parameters and FLOPs, thereby avoiding the computational burden and representation conflicts of full-image restoration. Therefore, the objective of LiDSeg is to augment conventional real-time semantic segmentation networks that are optimized on clean-weather accuracy for weather robustness. The main contributions of this work are as follows.

  1. We introduce LiDSeg, an end-to-end framework that augments a real-time semantic segmentation network with a lightweight linear-attention guidance to improve robustness in adverse weather conditions, without full-image reconstruction at inference.

  1. We utilize an asymmetric training strategy in which auxiliary image reconstruction supervises the guidance encoder to learn features robust to weather corruptions, and the reconstruction decoder is discarded at inference to preserve deployment efficiency while retaining weather-robust guidance.

  1. We design a Lightweight Gated Fusion (LGF) module that selectively integrates multiscale guidance into the semantic segmentation networks, preserving clear-weather accuracy while enhancing robustness in fog, rain, snow, and night.

LiDSeg was evaluated on multiple benchmarks. On Foggy Cityscapes, guided variants reach 74.1–74.3% dense-fog mIoU within 0.5–0.7 below FreD-ViT (109M) at 9–21× fewer parameters. On ACDC, LiDSeg shows similar robustness across several adverse weather conditions.

Related Works

Balancing segmentation accuracy and efficiency is critical for resource-constrained devices. Although multibranch architectures, such as BiSeNetV2 [56], PIDNet [53], DDRNet [14], STDC [10], and transformer-based approaches, such as TopFormer [58], RTLinearFormer [3], and SeaFormer [48], achieve efficient feature extraction on clear-weather datasets, they lack explicit mechanisms to handle feature degradation [8] in adverse weather. The direct application of these standard segmentation models to adverse conditions leads to significant mIoU drops [32, 56]. For instance, the BiSeNetV2 mIoU falls from 69.1% on clear Cityscapes to 54.2% on dense foggy Cityscapes (β=0.02\beta=0.02), and the DDRNet mIoU falls from 77.8% to 56.1%. Similar degradation occurs in ACDC and zero-shot datasets. Even robust conventional semantic networks, such as DeepLabV3+ [4] and HRNet [50] struggle despite their large model size [2, 59]. On the other hand, although large transformers, such as Mask2Former [6] enhance robustness, their expensive computational requirements restrict their applicability to resource-constrained devices. Recent weather-aware aggregation [36] and language-guided approaches [49] have also added complexity without fully resolving domain shifts. Overall, to address the challenge of semantic segmentation in adverse weather conditions, current methods typically fall into two main categories.

Restore-then-Segment. Sequential pipelines, such as TransWeather [46] and WeatherStream [59] treat restoration as a preprocessing step. Models, such as Restormer [57], DehazeFormer [43], and LDFormer [37] employ attention mechanisms for high-fidelity restoration before segmentation. However, this approach introduces significant latency and prioritizes pixel fidelity over semantic robustness [2]. Restoration often results in oversmoothed artifacts that obscure fine details. In addition, the disjoint sequential optimization of restoration and segmentation causes error accumulation [37, 44]. LiDSeg differs from sequential restoration-then-segment and restoration–segmentation multitasking by using dehazing only as lightweight training-time guidance, with no restoration at inference. It linearizes attention in the dehazing encoder and uses reconstruction solely to supervise the guidance branch, reducing computation and avoiding restoration-induced error accumulation. LGF then selectively integrates multiscale guidance features instead of using generic concatenation or U-Net-style fusion.

Domain Adaptation. Alternatively, domain-adaptation techniques bridge the feature gap through style transfer [19, 24, 36] and adversarial learning [47, 55]. Frequency-based methods, such as FreD [7], also attempt to separate style from structure and are trained using unsupervised domain adaptation (UDA). However, UDA approaches, such as Cum-Former [29], often involve high training complexity for marginal improvements in accuracy. Furthermore, although recent frameworks that leverage temporal correlation [23] and generalizable systems [54] have improved adaptation, they remain limited by computational inefficiencies. On the other hand, recent adverse-weather segmentation models, such as (HRDA [16], MIC [17], HALO [13], CoDA [28], WA2{}^{2}Net [36]) rely on >85>85M backbones and report no edge latency, whereas real-time architectures (BiSeNetV2, DDRNet, STDC) lack explicit weather adaptation. LiDSeg co-designs reconstruction supervision and channel gating for image degradations in resource-constrained environments, where reconstruction is used as an auxiliary task to penalize weather-specific encodings, and LGF modulates guidance with minimal overhead.

Proposed Methodology

LiDSeg Problem Formulation and Overview

Adverse weather conditions degrade image quality and corrupt semantic features essential for accurate scene understanding. This degradation can be mathematically formalized as:

I=D(J;θ),(1)\mathbf{I} = \mathcal{D}(\mathbf{J}; \boldsymbol{\theta}), \tag*{(1)}

where J\mathbf{J} denotes the clean image, I\mathbf{I} represents the observed degraded image, and D(⋅)\mathcal{D}(\cdot) is a degradation parameterized by condition-specific degradation variables θ\boldsymbol{\theta} (encompassing rain, snow, fog, and nighttime illumination). Depending on the adverse condition, D\mathcal{D} exhibits distinct yet structurally related physical degradation, from sparse additive occlusions (e.g., rain streaks [12]) to dense, depth-dependent multiplicative attenuation and global illumination changes [30, 41]. These conditions simultaneously degrade local high-frequency details and global contextual cues [25]. Across fog, rain, and snow, much of the corruption is channel-coherent, which motivates channel-wise fusion rather than heavy spatial attention; night is the partial exception (§ 4.5), where this assumption weakens.

Because semantic annotations depend on static scene content, which remains unaffected by weather, the objective is to learn f(⋅)f(\cdot) that is robust to weather-induced degradations:

f(D(J;θ))≈f(J),∀θ.(2)f(\mathcal{D}(\mathbf{J}; \boldsymbol{\theta})) \approx f(\mathbf{J}), \quad\forall\boldsymbol{\theta}. \tag*{(2)}

To enforce this robustness, LiDSeg introduces a dual-branch architecture (Figure 1), where the segmentation branch uses lightweight models [10, 14, 53, 56] and the guidance branch uses a linearized encoder inspired by DehazeFormer [43] to extract multi-scale features. LGF modules (Eq. (9)) integrate guidance into the segmentation branch with minimal overhead.

Guidance Encoder Architecture

The guidance encoder extracts multiscale features to support the robustness objective in Eq. (2). It produces feature maps at resolutions 1/81/8, 1/161/16, and 1/321/32 with channel dimensions {32,64,128}\{32,64,128\}, matching the segmentation branch to enable direct fusion without additional projection or resampling layers.

Processing high resolution images using resource constrained devices requires efficient attention mechanisms. Standard self-attention uses softmax(QKTQK^{\mathsf{T}})VV with O(N2)\mathcal{O}(N^2) complexity, making it difficult to process high-resolution features on resource-constrained devices. To address this, we adopt a linear-attention formulation inspired by [21], which exploits matrix associativity to compute the key-value product:

Attn⁡(Q,K,V)=Q(ρ(K)TV),(3)\operatorname{Attn}(\mathbf{Q},\mathbf{K},\mathbf{V}) = \mathbf{Q}\left(\rho(\mathbf{K})^{\mathsf{T}}\mathbf{V}\right), \tag*{(3)}

where Q,K,V∈RN×D\mathbf{Q},\mathbf{K},\mathbf{V} \in\mathbb{R}^{N \times D} are spatially flattened features, and ρ(⋅)\rho(\cdot) denotes softmax along the spatial dimension. The key-value product ρ(K)⊤V∈RD×D\rho(\mathbf{K})^{\top}\mathbf{V} \in\mathbb{R}^{D \times D} produces a compact D×DD \times D context matrix rather than an N×NN \times N attention matrix, reducing complexity to O(N⋅D2)O(N \cdot D^{2}).

To limit the propagation of weather artifacts through global feature aggregation, we add a learnable gate inspired by Gated Linear Units [9, 42]:

GLA⁡(x)=(Attn⁡(Q,K,V)⊙ϕ(G))WO,(4)\operatorname{GLA}(\mathbf{x}) = \left(\operatorname{Attn}(\mathbf{Q},\mathbf{K},\mathbf{V}) \odot\phi(\mathbf{G})\right)\mathbf{W}_{O}, \tag*{(4)}

where x∈RN×D\mathbf{x} \in\mathbb{R}^{N \times D} denotes the input feature map and WO∈RD×D\mathbf{W}_{O} \in\mathbb{R}^{D \times D} is the learnable output projection. The attention components are computed via linear transformations: Q=xWQ\mathbf{Q} = \mathbf{x}\mathbf{W}_{Q}, K=xWK\mathbf{K} = \mathbf{x}\mathbf{W}_{K}, and V=xWV\mathbf{V} = \mathbf{x}\mathbf{W}_{V}. The gating signal G∈RN×D\mathbf{G} \in\mathbb{R}^{N \times D} is derived from the input as G=xWG\mathbf{G} = \mathbf{x}\mathbf{W}_{G}, where WQ,WK,WV,WG∈RD×D\mathbf{W}_{Q},\mathbf{W}_{K},\mathbf{W}_{V},\mathbf{W}_{G} \in\mathbb{R}^{D \times D} are learned weights, and ϕ(⋅)\phi(\cdot) is the SiLU activation. This generates a soft mask that suppresses weather-artifact responses across channels through element-wise multiplication (⊙\odot) before the features are projected by WO\mathbf{W}_{O}.

Although attention captures global context, it lacks explicit modeling of local spatial patterns, such as edges and fine textures [52]. To complement the global receptive field of GLA, each encoder block incorporates a Gated-Dconv Feed-Forward Network (GDFN) [57] with depth-wise convolutions for local spatial mixing:

x^(l)=x(l−1)+GLA⁡(LN⁡(x(l−1))),\hat{\mathbf{x}}^{(l)} = \mathbf{x}^{(l-1)} + \operatorname{GLA}\left(\operatorname{LN}\left(\mathbf{x}^{(l-1)}\right)\right),
x(l)=x^(l)+GDFN⁡(LN⁡(x^(l))),(5)\mathbf{x}^{(l)} = \hat{\mathbf{x}}^{(l)} + \operatorname{GDFN}\left(\operatorname{LN}\left(\hat{\mathbf{x}}^{(l)}\right)\right), \tag*{(5)}
[x1,x2]=DWConv⁡(Win(x)),GDFN⁡(x)=Wout(Gate⁡(x1)⊙x2),(6)[\mathbf{x}_{1},\mathbf{x}_{2}] = \operatorname{DWConv}\left(\mathbf{W}_{\mathrm{in}}(\mathbf{x})\right), \qquad\operatorname{GDFN}(\mathbf{x}) = \mathbf{W}_{\mathrm{out}}\left(\operatorname{Gate}(\mathbf{x}_{1}) \odot\mathbf{x}_{2}\right), \tag*{(6)}

where LN is Layer Normalization, DWConv is depth-wise convolution, Gate(⋅)(\cdot) denotes the gating activation, and Win\mathbf{W}_{\mathrm{in}} and Wout\mathbf{W}_{\mathrm{out}} are linear projections. Together, GLA and GDFN enable the encoder to learn features that are both globally coherent and locally precise, which is essential for robust segmentation under weather degradation.

Auxiliary Reconstruction Decoder

The guidance encoder learns weather-robust representations using an auxiliary reconstruction decoder. This decoder follows a U-Net-like architecture with Restormer-style refinement [57]. At each decoder stage i∈{2,1}i \in\{2,1\},

di=Φi(Wireduce[Up⁡(di+1)∥ei]),(7)\mathbf{d}_{i} = \Phi_{i}\left(\mathbf{W}_{i}^{\mathrm{reduce}}\left[\operatorname{Up}(\mathbf{d}_{i+1}) \mathbin{\|} \mathbf{e}_{i}\right]\right), \tag*{(7)}

where di,ei∈RCi×Hi×Wi\mathbf{d}_{i},\mathbf{e}_{i} \in\mathbb{R}^{C_{i} \times H_{i} \times W_{i}} denote the decoder and encoder features at stage ii with channel dimension CiC_{i}, Up⁡(⋅)\operatorname{Up}(\cdot) is PixelShuffle with scale factor 2, [⋅∥⋅][\cdot\mathbin{\|} \cdot] denotes channel-wise concatenation, Wireduce∈RCi×2Ci\mathbf{W}_{i}^{\mathrm{reduce}} \in\mathbb{R}^{C_{i} \times2C_{i}} is a 1×11 \times1 convolution reducing channel from 2Ci2C_{i} to CiC_{i}, and Φi(⋅)\Phi_{i}(\cdot) denotes a stack of Restormer-like blocks (channel-wise attention and GDFN) for feature refinement. The decoder output passes through a progressive upsample that applies three successive 2×2 \times PixelShuffle operations (from 1/8 to full resolution) with intermediate refinement. The final reconstruction employs global residual learning as follows:

xrec=xin+Δx,(8)\mathbf{x}_{\mathrm{rec}} = \mathbf{x}_{\mathrm{in}} + \Delta\mathbf{x}, \tag*{(8)}

where xin∈R3×H×W\mathbf{x}_{\mathrm{in}} \in\mathbb{R}^{3 \times H \times W} is the degraded image and Δx∈R3×H×W\Delta\mathbf{x} \in\mathbb{R}^{3 \times H \times W} is the residual predicted by the decoder. The auxiliary decoder adds ∼15%\sim15\% peak GPU memory during training only.

Lightweight Gated Fusion (LGF)

Given the spatial and condition dependent nonuniformity of weather degradation, we propose an LGF that selectively fuses guidance features into segmentation features at multiple stages using a channel-wise gating mechanism. At fusion stages k∈{3,4,5}k \in\{3,4,5\}, the fused features are computed as

fk′=fk+λk⋅σ(W2δ(W1GAP⁡([fk,gk])))⊙gk,(9)\mathbf{f}'_k = \mathbf{f}_k + \lambda_k \cdot\sigma\left(W_2 \delta\left(W_1 \operatorname{GAP}([\mathbf{f}_k,\mathbf{g}_k])\right)\right) \odot\mathbf{g}_k, \tag*{(9)}

where fk,gk∈RCk×Hk×Wk\mathbf{f}_k,\mathbf{g}_k \in\mathbb{R}^{C_k \times H_k \times W_k} denote semantic and guidance features (Ck∈{32,64,128}C_k \in\{32,64,128\}), respectively. Leveraging the squeeze-and-excitation principle [18], we concatenate the inputs and apply Global Average Pooling (GAP) to extract global context, followed by a bottleneck MLP (reduction r=8r=8, weights W1,2W_{1,2}) with ReLU δ\delta and sigmoid σ\sigma to generate channel-wise gates. Here λk\lambda_k is a learnable per-stage scalar that controls how much guidance is needed at each scale; the “λ×\lambda\times” in Figure 1 denotes this learned gate. Weather corruption in fog, rain, and snow is predominantly channel-coherent (Eq. (1)); we verify this by measuring the cross-channel correlation of the clear→\todegraded residual, which is high for fog/rain/snow (ρˉ>0.8\bar{\rho}>0.8) but lower for night (ρˉ≈0.5\bar{\rho}\approx0.5, consistent with § 4.5). Channel gating targets this structure with far fewer parameters and FLOPs, whereas spatial or cross-attention fusion costs more fusion parameters for only a marginal accuracy gain (Table 8). These gates dynamically modulate gk\mathbf{g}_k, preserving clear-weather accuracy while enhancing robustness in adverse conditions.

GuidanceAttn.Struct.FusionRec.Params (M)FLOPs (G)Mean
Dehazing + SegmentSoftmaxSeq.–Yes6.7631.467.8
Heavy guidanceSoftmaxAsy.ConcatYes5.9446.267.1
Heavy guidanceSoftmaxAsy.LGFYes5.9446.075.4
LiDSegLinearSeq.–Yes6.3143.066.3
LiDSegLinearAsy.ConcatYes5.7110.866.1
LiDSegLinearAsy.SpatialYes6.3126.575.6
LiDSegLinearAsy.CrossYes6.4131.575.9
LiDSegLinearAsy.LGFYes5.22108.875.4

Table 8. Architectural design control on Foggy Cityscapes val using BiSeNetV2. Mean denotes mIoU averaged over β∈{0.005,0.01,0.02}\beta\in\{0.005, 0.01, 0.02\}. Reconstruction supervision is retained for all guided variants. Params and FLOPs correspond to the deployed networks.

Loss Function

The training loss combines segmentation and reconstruction to achieve a balance between the segmentation focus on discriminative features and the reconstruction need for pixel-level accuracy. Therefore, the total loss Ltotal\mathcal{L}_{\mathrm{total}} is computed using the task-dependent uncertainty weight [22] as shown in Eq. (10).

Ltotal=1σseg2Lseg+1σrec2Lrec+log⁡σseg2+log⁡σrec2,(10)\mathcal{L}_{\mathrm{total}} = \frac{1}{\sigma_{\mathrm{seg}}^2}\mathcal{L}_{\mathrm{seg}} + \frac{1}{\sigma_{\mathrm{rec}}^2}\mathcal{L}_{\mathrm{rec}} + \log\sigma_{\mathrm{seg}}^2 + \log\sigma_{\mathrm{rec}}^2, \tag*{(10)}

where σseg2\sigma_{\mathrm{seg}}^2 and σrec2\sigma_{\mathrm{rec}}^2 are learnable task uncertainties. Higher uncertainty downweights the corresponding loss, whereas the log regularizes to prevent the trivial solution σ→∞\sigma\to\infty.

Segmentation loss. Lseg\mathcal{L}_{\mathrm{seg}} combines cross-entropy, Lovász-Softmax [1], and auxiliary supervision:

Lseg=LCE+λLLLov+λauxLaux.(11)\mathcal{L}_{\mathrm{seg}} = \mathcal{L}_{\mathrm{CE}} + \lambda_L \mathcal{L}_{\mathrm{Lov}} + \lambda_{\mathrm{aux}}\mathcal{L}_{\mathrm{aux}}. \tag*{(11)}

We set λL=0.5\lambda_L = 0.5 and λaux=0.4\lambda_{\mathrm{aux}} = 0.4. These values follow established practice: λaux=0.4\lambda_{\mathrm{aux}} = 0.4 matches the deep-supervision weighting in PSPNet [60] and BiSeNetV2 [56], ensuring that auxiliary heads do not dominate gradients. The weight λL=0.5\lambda_L = 0.5 balances the IoU-based Lovász optimization against pixel-wise cross-entropy, as recommended in [1]. A sensitivity sweep (in supplementary) shows LiDSeg is robust to this choice.

Reconstruction loss. For reconstruction, the Charbonnier loss is used for training.

Lrec=1N∑n=1N(x^n−xnclear)2+ε2,ε=10−3,(12)\mathcal{L}_{\mathrm{rec}} = \frac{1}{N}\sum_{n=1}^{N}\sqrt{(\hat{x}_n-x_n^{\mathrm{clear}})^2+\varepsilon^2}, \qquad\varepsilon=10^{-3}, \tag*{(12)}

where NN is the number of pixels, x^n\hat{x}_n is the nn-th pixel of the reconstructed image, xnclearx_n^{\mathrm{clear}} is the clear ground-truth pixel, ε\varepsilon is used for numerical stability. The adverse-to-clear mapping is not assumed to be uniquely identifiable because weather can obscure scene information [31]. LiDSeg therefore uses LrecL_{\mathrm{rec}} only as a pretraining regularizer for degradation-stable guidance features rather than requiring a unique photorealistic reconstruction at deployment; the decoder is discarded afterward, and § 6.1 isolates the contribution of this supervision.

Experimental Results and Discussion

Experimental Setup

We implemented LiDSeg in PyTorch and trained on NVIDIA Tesla V100 GPUs using the AdamW optimizer (lr=10−3\mathrm{lr}=10^{-3}, weight decay 10−410^{-4}), automatic mixed precision, and a cosine learning rate schedule. During paired Foggy Cityscapes pretraining, the guidance encoder is optimized with both segmentation and reconstruction objectives. For ACDC fine-tuning, the reconstruction decoder and LrecL_{\mathrm{rec}} are removed, and the segmentation branch, LGF, and pretrained guidance encoder are fine-tuned using segmentation labels only. We evaluate both Uber configurations (optimized across all adverse weather conditions simultaneously) and condition-specific Expert models (trained on a single weather condition each). Inference latency is measured in frames per second (FPS) on an NVIDIA Jetson AGX Orin.

Datasets. In reality, capturing perfectly aligned pairs of clear and adverse weather image for training is nearly impossible because the same dynamic scene (with moving objects, vehicles, and pedestrians) cannot be recorded at the exact same timestamp under both clear and adverse weather without any other variations, making reliable ground-truth annotation extremely difficult. Therefore, we rely on synthetic pairs that provide pixel-level alignment between clean and weather-affected images while isolating only the weather effects to learn robust features. Given this reality, we pretrain on Foggy Cityscapes [41], which provides synthetic paired clear and foggy images across a progressive curriculum of attenuation coefficients β∈{0.005,0.01,0.02}\beta\in\{0.005,0.01,0.02\}. This also allows direct comparison with recent works on adverse weather conditions such as FreD [2], IBNet [34], and IterNorm [20]. Real-world evaluation and target-domain fine-tuning (which requires no clear image pairs) utilize ACDC [39], spanning real fog, night, rain, and snow. Moreover, for evaluating generalization, we assessed guided models via zero-shot on Foggy Zurich v2 [40] and Foggy Driving [41].

Metrics. We report mIoU per class and per condition. For ACDC, Pooled All denotes the benchmark’s class-mean IoU over the pooled all-condition set and is not the arithmetic mean of Fog, Night, Rain, and Snow. To analyse the correlation between reconstruction and segmentation, we also track Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM).

Learned Feature Robustness and Training Dynamics

To evaluate model robustness, we analyze feature stability by quantifying the L2L_2 feature drift between corresponding clear and foggy scenes. As illustrated in the Feature Drift Distribution (Figure 2a), LiDSeg (blue) exhibits a narrow distribution clustered near zero, demonstrating better stability against weather-induced domain shifts compared to the baseline (pink). Quantitatively, LiDSeg substantially reduces the mean feature drift to μ=0.186\mu= 0.186 relative to the baseline (μ=0.343\mu= 0.343). This consistency is validated by per-scene comparisons (Figure 2b), where LiDSeg yields uniformly lower drift across all test scenes.

Figure 2. Feature stability on Foggy Cityscapes val (BiSeNetV2) (β=0.02\beta=0.02): (a) L2L_2 feature drift (LiDSeg μ=0.186\mu=0.186 vs. baseline μ=0.343\mu=0.343). (b) Per-scene drift (LiDSeg lower on all scenes).

Performance and stability on Foggy Cityscapes

Figure 3 plots validation mIoU and guidance-encoder reconstruction PSNR for all four guided models over the fog curriculum (β=0.005→0.01→0.02\beta=0.005\to0.01\to0.02). Reconstruction quality and segmentation accuracy co-evolve throughout training, where both metrics slightly dip at each denser fog in a synchronized manner. At each transition, the encoder encounters a sudden distribution shift because denser fog attenuates high-frequency content, reducing reconstruction quality (lower PSNR) and perturbing the segmentation feature space (lower mIoU). However this shift is relatively small compared to baselines as illustrated in Table 1. Additionally, the subsequent rise in both curves is consistent with the recovery of latent structure useful for segmentation. Those patterns are nearly identical across all LiDSeg variants despite architectural differences (bilateral fusion, dual-resolution, PID, STDC). We also compared LiDSeg-guided models with adverse-weather and baseline models. Table 1 reports the Foggy Cityscapes validation

Figure 3. Reconstruction and segmentation training dynamics on Foggy Cityscapes validation set (four guided models; fog curriculum β=0.005→0.01→0.02\beta=0.005\to0.01\to0.02). Validation mIoU (solid, left axis) and guidance-encoder reconstruction PSNR (dashed, right axis).

MethodBackboneParamsGFLOPsClear (mIoU)Light (mIoU)Medium (mIoU)Dense (mIoU)Δ\Delta
Real-time semantic segmentation models
BiSeNetV2 [56]–3.36M99.769.160.958.854.2↓\downarrow14.9
DDRNet-23s [14]–5.70M36.777.866.162.256.1↓\downarrow21.7
STDC1-Seg [10]–8.43M166.371.867.562.556.1↓\downarrow15.7
PIDNet-L [53]–37.31M276.274.068.463.058.9↓\downarrow15.1
Unguided models with incremental (matched budget, same curriculum)
BiSeNetV2-wide†BiSeNetV25.22M108.870.464.260.656.9↓\downarrow13.5
DDRNet-23s-wide†DDRNet7.65M46.078.268.564.057.8↓\downarrow20.4
Models use heavy backbones
IBNet [34]ResNet-50>25.6M>10076.572.467.959.5↓\downarrow17.0
IterNorm [20]ResNet-50>25.6M>10076.072.068.360.7↓\downarrow15.3
SW [35]ResNet-50>25.6M>10077.073.369.461.7↓\downarrow15.3
ISW [7]ResNet-50>25.6M>10076.272.167.960.1↓\downarrow16.1
DeepLabV3+ [4]ResNet-10162.7M203279.667.165.261.6↓\downarrow18.0
SegFormer [52]MiT-B464.1M124081.070.566.162.3↓\downarrow18.7
Mask2Former [6]Swin-B107M46683.576.274.570.8↓\downarrow12.7
FreD-ViT [2]Swin-B109M46682.879.578.074.8↓\downarrow8.0
LiDSeg (BiSeNetV2)–5.22M108.876.776.375.774.2↓\downarrow2.5
LiDSeg (DDRNet-23s)–7.65M46.076.976.275.674.1↓\downarrow2.8
LiDSeg (STDC1-Seg)–11.48M176.376.976.275.574.3↓\downarrow2.6
LiDSeg (PIDNet-L)–39.54M285.976.876.275.574.2↓\downarrow2.6

Table 1. Fog density robustness on Foggy Cityscapes val (mIoU %). Clear is clear-weather Cityscapes mIoU; Δ\Delta is the clear→\rightarrowdense (β=0.02\beta=0.02) mIoU drop (smaller is better). GFLOPs at 1024×20481024 \times2048. †: denotes incremental widening to match the guided parameter/FLOP budget and trained on an identical fog curriculum without the guidance branch.

mIoU at β∈{0.005,0.01,0.02}\beta\in\{0.005, 0.01, 0.02\}. LiDSeg rows are taken from the per-stage training of Figure 3. Every guided model degrades by only 2.5–2.8 mIoU points from clear to dense fog. This drop is less than half that of FreD-ViT and substantially smaller than those of the unguided baselines. The scaled unguided models (BiSeNetV2-wide and DDRNet-23s-wide), widened to 5.2/7.7M budget and trained on the identical curriculum without the guidance branch, reach only 56.9/57.8% dense-fog mIoU. Therefore, scaling the model to match the parameter and FLOP budget recovers at most +2.7 mIoU points over the vanilla model, whereas LiDSeg recovers +20.0 points. This gain results from the reconstruction-guided robustness rather than added capacity. On dense-fog accuracy LiDSeg reaches 74.1–74.3%, within 0.5–0.7 mIoU points of FreD-ViT (74.8%) while using substantially fewer parameters, placing it on the robustness and efficiency frontier of adverse weather condition semantic segmentation design.

LiDSeg Guidance Performance on ACDC

The objective of LiDSeg is to learn weather-robust features from synthetic paired images and transfer them to real-world datasets such as ACDC. During ACDC fine-tuning, no paired clear target is used: the reconstruction decoder and Lrec\mathcal{L}_{rec} are removed, and the segmentation branch, LGF, and guidance encoder are updated with ACDC segmentation labels only. Because supervised fine-tuning and UDA use different target-domain supervision, Table 2 labels each protocol explicitly; UDA rows are included as task-level references for comparisons. LiDSeg-PIDNet obtains 75.07% mIoU, comparable with supervised HRNet-W48 (75.0% mIoU, 696 GFLOPs) and other heavy UDA models, while night remains the most challenging condition.

MethodProtocolParamsGFLOPsFogNightRainSnowPooled AllMean
Supervised ACDC labels
RefineNet [27]Sup.118M–65.755.568.765.965.364.0
DeepLabV2 [39]Sup.44M–54.545.359.357.155.354.1
DeepLabV3+ [4]Sup.62.7M203269.160.974.169.670.068.4
HRNet-W48 [50]Sup.65.9M69674.765.377.776.375.073.5
Mask2Former [6]Sup.107M466––––77.3–
ViT-Adapter [5]Sup.>320M>1000––––78.4–
UDA task-level references
DANNet [51]UDA––––––50.0–
DAFormer [15]UDA84.0M–––––55.4–
HRDA [16]UDA85.7M251––––68.0–
MIC [17]UDA85.7M251––––70.4–
HALO [13]UDA86.0M259––––71.9–
CoDA [28]UDA90.5M271––––72.6–
WA²Net [36]UDA>85M>255––––73.8–
LiDSeg (PIDNet-L)Sup.-FT39.54M285.973.9964.5878.5877.1875.0773.58

Table 2. ACDC test mIoU (%). Pooled All is the benchmark pooled all-condition evaluation; Mean is the arithmetic mean of Fog, Night, Rain, and Snow where per-condition scores are available. Sup.: supervised ACDC labels; Sup.-FT: LiDSeg fine-tuned with ACDC segmentation labels after synthetic paired pretraining; UDA: adaptation with unlabeled target images. (–): not reported.

ACDC: Per-Condition Per-Class performance

Across all conditions, Uber models consistently outperform condition-specific Experts, showing that joint training across adverse conditions strengthens the shared representation. Cross-condition evaluation in Table S8 further supports this finding, with strong transfer among the channel-coherent fog, rain, and snow conditions. Table 4 reports the corresponding class-wise IoU results. Fog remains the strongest setting, consistent with Foggy-Cityscapes pretraining, whereas night remains the most challenging condition.

Cond.SetupModelroadsidew.build.wallfencepolelightsignveget.terrainskypersonridercartruckbustrainmotorc.bicyclemIoU
FogUberBiSeNetV297.185.988.869.944.868.078.075.090.661.198.368.277.585.564.396.291.753.962.176.7
DDRNet97.285.588.570.344.968.177.374.290.560.798.366.975.184.562.295.790.552.165.176.2
STDC97.085.488.570.545.068.376.674.390.461.198.367.776.485.266.595.790.857.865.676.9
PIDNet97.286.188.769.944.368.678.375.190.661.898.369.674.985.263.295.891.652.365.676.7
ExpertBiSeNetV297.286.687.568.041.664.672.467.990.261.498.361.067.285.054.486.287.631.257.371.9
DDRNet96.584.786.266.340.061.370.065.689.861.598.261.868.383.853.288.383.630.658.671.0
STDC97.285.687.768.342.063.372.468.290.361.598.362.768.084.553.286.083.433.957.171.8
PIDNet96.383.785.767.537.859.069.766.089.458.298.260.365.383.348.288.981.030.557.069.8
NightUberBiSeNetV293.872.776.842.342.255.467.244.269.734.680.241.230.471.931.638.385.032.947.055.7
DDRNet93.973.476.142.140.754.267.241.268.335.378.538.732.569.530.138.681.734.945.554.9
STDC93.973.776.542.741.254.464.941.868.835.378.840.631.169.930.238.582.834.247.155.1
PIDNet94.173.677.043.542.155.367.444.269.834.980.239.232.371.631.938.585.831.450.756.0
ExpertBiSeNetV292.470.173.438.642.054.661.839.666.534.776.935.931.668.433.835.372.934.945.353.1
DDRNet87.954.059.431.630.648.836.930.855.436.862.530.935.740.234.937.531.937.731.842.9
STDC88.056.863.430.930.249.845.230.356.036.662.730.635.845.035.236.439.937.630.544.3
PIDNet91.768.872.535.439.753.756.835.964.735.174.035.034.465.934.834.276.735.634.151.5
RainUberBiSeNetV292.272.688.140.843.360.075.576.592.766.098.467.739.289.858.592.077.047.338.469.3
DDRNet92.471.988.340.648.960.375.576.092.566.598.467.042.289.859.492.376.039.638.869.3
STDC92.371.688.341.446.159.975.176.092.666.598.468.151.890.156.493.576.340.544.370.0
PIDNet92.572.788.441.446.260.375.476.492.866.198.567.450.890.260.793.276.340.741.670.1
ExpertBiSeNetV292.371.986.841.840.259.374.774.392.968.198.467.742.889.250.089.771.0……68.4
DDRNet91.969.986.141.340.758.474.573.492.667.598.466.436.388.647.089.769.342.337.566.9
STDC91.368.986.140.341.757.073.872.192.767.598.466.730.388.144.889.864.237.837.865.8
PIDNet92.070.585.940.440.057.674.772.992.667.498.367.234.088.647.489.167.945.238.766.9
SnowUberBiSeNetV293.377.187.264.145.262.582.076.789.354.697.872.034.288.565.182.592.756.760.672.7
DDRNet92.173.586.962.244.560.081.976.589.055.097.969.935.188.068.781.193.558.361.072.4
STDC92.274.386.961.744.260.681.775.989.055.397.869.835.287.970.677.092.857.360.672.1
PIDNet93.176.387.465.244.862.182.177.389.454.197.974.235.888.572.081.893.264.362.873.8
ExpertBiSeNetV290.970.484.760.744.956.578.274.287.849.997.768.435.484.541.249.886.755.658.267.1
DDRNet91.772.284.058.044.955.377.374.287.947.097.566.034.684.940.141.788.754.356.966.2
STDC91.974.085.860.946.557.779.276.188.751.497.772.031.685.836.553.086.758.558.968.0
PIDNet89.667.183.154.838.154.376.373.987.243.397.559.837.383.533.435.682.452.253.763.3

Table 4. LiDSeg guided ACDC val per-condition IoU (%). Gray rows = Uber; white rows = Expert.

ModelExpertFogNightRainSnow
BiSeNetV2 (5.22M, 108.8 GFLOPs)
BiSeNetV2Fog71.941.7<u>54.4</u>51.5
BiSeNetV2Night50.453.143.442.1
BiSeNetV2Rain53.342.868.454.2
BiSeNetV2Snow51.840.757.167.1
DDRNet-23s (7.65M, 46.0 GFLOPs)
DDRNet-23sFog71.037.252.750.3
DDRNet-23sNight47.842.942.041.0
DDRNet-23sRain50.638.066.9<u>53.0</u>
DDRNet-23sSnow49.236.355.466.2
STDC1-Seg (11.48M, 176.3 GFLOPs)
STDC1-SegFog71.838.051.852.7
STDC1-SegNight48.844.341.343.2
STDC1-SegRain51.738.965.855.4
STDC1-SegSnow50.337.1<u>54.5</u>68.0
PIDNet-L (39.54M, 285.9 GFLOPs)
PIDNet-LFog69.840.3<u>51.7</u>46.6
PIDNet-LNight46.551.541.339.3
PIDNet-LRain49.341.366.949.1
PIDNet-LSnow47.939.354.463.3

Table S8. ACDC val mIoU (%) for condition-specific Expert cross-evaluation. Rows: training condition; columns: test condition. Diagonal: same-condition performance. Bold and <u>underline</u> indicate the best and second-best off-diagonal values, respectively, for each model.

MethodRoadSidewalkBuildingWallFencePoleLightSignVegetationTerrainSkyPersonRiderCarTruckBusTrainMotorcycleBicyclemIoU
ADVENT [47]72.914.340.516.621.29.317.421.263.823.818.332.619.569.536.234.546.226.936.132.7
MGCDA† [38]73.428.769.919.326.336.853.053.375.432.084.651.026.177.643.245.953.932.741.548.7
DANNet† [51]84.354.277.638.030.018.941.635.271.339.486.648.729.276.241.643.058.632.643.950.0
DAFormer [15]58.451.384.042.735.150.730.057.074.852.851.358.332.682.758.354.982.444.150.755.4
SSAC† [?]76.559.882.752.634.551.258.161.263.958.167.261.540.081.859.261.783.447.555.659.6
HRDA [16]88.357.988.155.236.756.362.965.374.257.785.968.845.788.576.482.487.752.760.468.0
MIC [17]90.867.189.254.540.557.262.068.476.361.887.071.349.489.775.786.889.156.963.070.4
HALO [13]94.279.888.260.251.164.178.265.687.955.795.566.320.788.982.289.387.950.459.071.9
CoDA [28]93.172.790.757.347.456.869.970.087.359.895.471.447.690.377.183.889.154.764.172.6
WA²Net [36]94.577.191.661.947.657.568.570.479.264.095.672.850.290.575.588.790.957.562.673.8
LiDSeg (PIDNet-L)94.9480.4591.5062.9556.0966.2576.7776.2489.7769.4996.5875.1954.9889.7464.9578.4688.5950.6062.7475.07

Table 3. All-conditions ACDC test per-class IoU (%). Comparison rows are reported by WA²Net [36]. LiDSeg (PIDNet-L) is Uber training on ACDC. Bold and underlined values denote the best and second-best results, respectively.

Why night remains the hardest condition. On ACDC-Night, the Uber models achieve 54.9–56.0% mIoU, which is below HRNet-W48 and represents the weakest relative to other conditions. This reflects a mismatch between the assumptions underlying channel-wise gating and nighttime degradation. Fog, rain, and snow primarily introduce atmospheric scattering and attenuation that are largely channel-coherent and approximately multiplicative (Eq. (1)), making them well suited to shared channel-wise modulation. In contrast, nighttime degradation is dominated by low illumination, signal-dependent sensor noise, non-uniform artificial lighting, and dynamic-range compression, which are not channel-coherent. Moreover, nighttime scenes contain limited recoverable high-frequency content, weakening the reconstruction-based auxiliary objective and reducing the robustness signal available during joint training. Consequently, all-condition optimization is driven more strongly by fog, rain, and snow than by night. Nevertheless, the shared weather-robust encoder still improves substantially over condition-specific Experts on nighttime scenes (up to 12.0 mIoU points), showing that cross-condition representation sharing remains beneficial (details in supplementary). Incorporating illumination-aware gating, e.g., conditioning LGF reduction on learned global luminance statistics, is a potential direction for future work.

Performance on Zero-Shot

LiDSeg is evaluated on Foggy Zurich v2 and Foggy Driving (Table 5). Once trained solely on synthetic fog (Foggy Cityscapes) and paired guidance, it generalizes to unseen real fog without any real-fog domain adaptation. Prior UDA, CDA, and STL methods leverage unlabeled Foggy Zurich images during training, whereas LiDSeg and the vanilla RefineNet baselines do not. Despite this stricter zero-shot protocol, LiDSeg outperforms all prior methods while using efficient real-time architectures, demonstrating that strong synthetic paired supervision alone can enable robust cross-domain transfer to real fog.

ModelBackboneMethodUnlabeled FZFZFD
RefineNet [27]ResNet-101SFT×34.644.3
RefineNet-lw [32]ResNet-101SFT×28.543.6
AdaptSegNet [45]ResNet-101UDA✓25.029.7
AdvEnt [47]ResNet-101UDA✓39.746.9
FDA [55]ResNet-101STL✓22.221.8
DANN [11]ResNet-101UDA✓43.146.0
CMAda2+ [40]ResNet-101CDA✓43.449.9
CMAda3+ [40]ResNet-101CDA✓46.849.8
FIFO [24]RefineNetSTL✓42.048.9
D2SL [44]RefineNet-lwSTL✓44.245.9
LiDSeg (BiSeNetV2)–Guided×53.550.4
LiDSeg (STDC1-Seg)–Guided×53.452.4

Table 5. Zero-shot mIoU (%) on Foggy Zurich v2 (FZ) and Foggy Driving (FD). SFT: supervised fine-tuning; UDA: unsupervised domain adaptation; STL: style transfer learning; CDA: curriculum domain adaptation. ✓: uses unlabeled real-fog images (Foggy Zurich) during training. No method uses FD during training.

Efficiency on Edge Hardware

Real-world deployment requires balancing accuracy with speed. However, prior work on adverse-weather semantic segmentation has focused almost exclusively on accuracy. Consequently, computational efficiency remains underexplored, and the absence of accuracy–efficiency trade-off baselines makes fair benchmarking in this domain difficult. We address this by decoupling the two factors: efficiency can be treated as an architectural property and evaluated against general-purpose real-time models, whereas robustness is evaluated against methods that explicitly target adverse weather. This separation is necessary because a direct comparison with general-purpose models would be unfair. Moreover, existing adverse-weather approaches typically gain robustness through computationally heavy designs, such as deep context aggregation, domain adaptation, or restoration preprocessing (e.g., dehazing before the downstream semantic task), all of which limit edge deployment. In contrast, LiDSeg decouples weather-robustness learning using a lightweight guidance module.

As shown in Table 6, our Jetson AGX Orin measurements confirm that on-device efficiency depends on more than FLOPs or parameter counts. For instance, Baseline DDRNet runs at 350.4 FPS despite having more parameters than BiSeNetV2 (5.70M vs. 3.36M). In contrast, SeaFormer-B has the lowest computational complexity (3.53 GFLOPs) yet reaches only 166.1 FPS, well below DDRNet’s 350.4 FPS at 9.17 GFLOPs. This indicates that hardware-level factors beyond raw FLOP counts also affect throughput. Adding LiDSeg guidance introduces only a modest latency overhead while preserving real-time throughput across all variants. LiDSeg(DDRNet) offers the strongest speed and robustness balance at 204.5 FPS, LiDSeg(STDC) achieves 188.0 FPS at 76.9% clear and 74.3% dense-fog mIoU, and LiDSeg(TopFormer-B) provides a middle ground at 126.4 FPS. LiDSeg stabilizes accuracy across fog densities, bounding the clear-to-dense-fog drop to Δ≤2.8\Delta\le2.8 mIoU points, whereas the unguided baselines degrade by up to 21.7 mIoU points. It is also more parameter-efficient, using 9×9 \times–21×21 \times fewer parameters than prior state-of-the-art methods such as FreD-ViT while achieving comparable or better robustness (Table 1). Together, these results provide LiDSeg with a favorable robustness–efficiency trade-off, where lightweight guidance delivers reliable adverse-weather performance without sacrificing the real-time, edge-deployable behavior of the underlying adapted model.

Table 6: Jetson AGX Orin efficiency at 512×1024. Guided = LiDSeg with the reconstruction decoder discarded at inference. Clear/Dense mIoU: Cityscapes clear-weather mIoU and Foggy Cityscapes dense-fog mIoU (β=0.02\beta=0.02); Δ\Delta is the clear→dense drop.

ModelVariantParams (M)GFLOPsFPSMem (MB)ClearDenseΔ\Delta
BiSeNetV2Baseline3.3624.9294.25169.154.2↓\downarrow14.9
DDRNetBaseline5.709.17350.44277.856.1↓\downarrow21.7
STDCBaseline8.4341.57267.04771.856.1↓\downarrow15.7
PIDNet-LBaseline37.3169.04116.911174.058.9↓\downarrow15.1
TopFormer-BBaseline7.069.69271.77675.859.0↓\downarrow16.8
SeaFormer-BBaseline8.583.53166.15877.764.6↓\downarrow13.0
LiDSeg (BiSeNetV2)Guided5.2227.2083.45876.774.2↓\downarrow2.5
LiDSeg (DDRNet)Guided7.6511.50204.55276.974.1↓\downarrow2.8
LiDSeg (STDC)Guided11.4844.08188.06976.974.3↓\downarrow2.6
LiDSeg (PIDNet-L)Guided39.5471.4892.613276.874.2↓\downarrow2.6
LiDSeg (TopFormer-B)Guided9.0112.02126.46376.974.2↓\downarrow2.7

Table 6.

Ablation Study

Reconstruction and fusion control. Table 7 isolates the effects of reconstruction supervision and LGF. Removing only LrecL_{\mathrm{rec}} while keeping the same guidance encoder and LGF reduces mean fog mIoU from 75.4% to 66.7% and ACDC mIoU from 73.6% to 67.0%, separating reconstruction supervision from added capacity. Reconstruction with concatenation reaches 66.1%, while LGF without reconstruction reaches 66.7%; combining both in full LiDSeg achieves 75.4%, showing their complementary contribution.

Table 7: Reconstruction and fusion control using BiSeNetV2. Mean denotes Foggy Cityscapes val mIoU averaged over β∈{0.005,0.01,0.02}\beta\in\{0.005,0.01,0.02\}; ACDC denotes validation pooled-All mIoU. Params and FLOPs correspond to the deployed networks.

SettingRec.FusionParams (M)FLOPs (G)MeanACDC
(a) Seg. onlyNo–3.3699.758.056.0
(b) Recon.+ConcatYesConcat5.70110.866.165.0
(c) LGF, no recon.NoLGF5.22108.866.767.0
(d) LiDSeg fullYesLGF5.22108.875.473.6

Table 7.

Architectural design control. Table 8 further ablates attention type, structural design, and fusion under fixed reconstruction supervision. Linear attention with asymmetric LGF matches Softmax accuracy at 75.4% mIoU while reducing computation from 446.0 to 108.8 GFLOPs. Sequential variants underperform while requiring more computation. Concatenation reduces mIoU to 66.1%. Spatial and cross-attention reach 75.6% and 75.9%, respectively, but at higher parameter and computational cost. Therefore, LGF provides near-best accuracy at the lowest cost, supporting the final linear-attention, asymmetric, gated design for edge deployment.

Model Scaling

The guidance encoder and LGF can be attached to segmentation networks that expose compatible s8/s16/s32 feature taps, adding a small overhead while improving robustness in adverse weather conditions. To test whether the gains transfer beyond the CNN-based real-time backbones, we apply the identical training protocol to two efficient transformer models, TopFormer-B [8] and SeaFormer-B [48] (Table 9). Uber training again outperforms the condition specific Experts, with fog remaining the strongest and night the weakest condition. This confirms that the guidance mechanism is not tailored to a particular architecture family but drives robustness gains across diverse semantic segmentation models.

BackboneSettingParams (M)GFLOPsFPSFogNightRainSnowMean
TopFormer-B [58]Baseline7.069.69271.767.545.061.563.559.4
Expert9.0112.02126.475.852.467.969.666.4
Uber76.455.369.571.468.2
SeaFormer-B [48]Baseline8.583.53166.167.044.561.063.058.9
Expert10.986.05130.375.452.067.569.266.0
Uber76.054.769.170.867.7

Table 9. Model-scaling: Transformer based semantic segmentation under the identical training protocol. Baseline is the unguided. Efficiency is reported at 512×1024512 \times1024 (Jetson AGX Orin).

Conclusion

We present LiDSeg, an efficient framework for robust semantic segmentation in adverse weather conditions. LiDSeg uses image reconstruction as an auxiliary task to supervise a lightweight guidance encoder and discards the reconstruction decoder at inference, improving the weather robustness of standard real-time semantic segmentation with minimal cost. Our experiments show that reconstruction-guided feature learning is more robust and efficient than using restoration as a preprocessor, while offering better efficiency than large model domain adaptation. LiDSeg maintains nearly stable accuracy from light to dense fog, transfers across diverse real-time models, generalizes zero-shot, and runs in real time on resource-constrained edge devices, supporting practical deployment. It effectively transfers features learned from synthetic degraded-clean image pairs but currently requires such pairs for reconstruction supervision, while weak or unpaired supervision remains unexplored. Nighttime remains a challenge because illumination weakens the channel-coherence assumption underlying our gating mechanism. Future work will explore efficient self-supervised variants without paired data, illumination-aware gating, temporal consistency for video segmentation, more adverse weather conditions, and a general asymmetric guidance paradigm for lightweight models and vision tasks affected by domain shift. Overall, LiDSeg improves weather robustness and efficiency through reconstruction-guided feature learning while remaining compact for edge deployment.

Acknowledgements

This work has received funding from the European Union’s Horizon Europe research and innovation programme under the Marie Skłodowska-Curie grant agreement No. 101169378, (TIRAMISU Doctoral Network), and from the Government of the Republic of Cyprus through the Cyprus Deputy Ministry of Research, Innovation and Digital Policy.

Supplementary Material

This supplement provides additional experimental results for the LiDSeg architecture and training strategy. Our fundamental premise is that the standard segmentation loss is insufficient to penalize weather-induced degradation. Therefore we use an auxiliary reconstruction loss with paired clear images to supervise the shared guidance encoder in learning degradation stable features, reducing L2L_2 feature drift from μ=0.343\mu= 0.343 (unguided) to μ=0.186\mu= 0.186 (guided). Because the reconstruction decoder is discarded at inference, LiDSeg avoids inference time restoration cost and restoration induced error propagation while retaining lightweight guidance, outperforming typical sequential restore-then-segment pipelines without requiring additional domain adaptation. Internally, the guidance encoder uses gated linear attention to achieve Softmax-level accuracy at approximately 4× fewer FLOPs while limiting the propagation of low-frequency artifacts by global linear kernels. Furthermore, because weather degradation is highly channel-coherent (ρˉ>0.8\bar{\rho} > 0.8 for fog, rain, and snow), our lightweight gated fusion (LGF) module selectively modulates guidance features with minimal overhead.

Comprehensive ablations confirm the stability and efficiency of our design choices. Sensitivity sweeps demonstrate that the model is robust to variations in loss weighting, where uncertainty-weighting outperforms equal-weighting. Additionally, the progressive fog training curriculum (β=0.005→0.01→0.02\beta= 0.005 \rightarrow0.01 \rightarrow0.02) is better than jointly training all densities, which drops the mean fog mIoU by a few points. Ablations also validate our multiscale fusion strategy as using all three scales consistently outperforms single and two scale configurations at strides 1/8, 1/16, and 1/32, confirming the benefit of multiscale guidance. Finally, comparisons against a sequential pipeline demonstrate that LiDSeg’s joint guidance yields better dense fog accuracy at a lower computational complexity.

Experimental Setup and Hyperparameters

We performed a thorough evaluation to ensure stable reproducibility. During paired Foggy Cityscapes pretraining, the uncertainty-weighted objective in Eq. (10) combines the segmentation loss with Charbonnier reconstruction loss. For ACDC fine-tuning, however, paired clear targets are not used: the reconstruction decoder and LrecL_{\mathrm{rec}} are removed, and we fine-tune the segmentation branch, LGF, and pretrained guidance encoder using segmentation labels only. We train one Uper model and four condition-specific Expert models per segmentation network. The segmentation loss integrates cross-entropy, Lovász-Softmax (λL=0.5\lambda_{\mathrm{L}} = 0.5), and auxiliary supervision (λaux=0.4\lambda_{\mathrm{aux}} = 0.4). The Foggy Cityscapes sensitivity analysis below evaluates this pretraining configuration; varying λaux\lambda_{\mathrm{aux}} shifts performance by up to 2.3 mIoU points depending on the model, consistently peaking at the λaux=0.4\lambda_{\mathrm{aux}} = 0.4 default (Table S1).

Semantic segmentation model0.20.30.4 (default)0.5
BiSeNetV274.174.375.475.2
DDRNet-23s73.074.275.374.1
STDC1-Seg75.175.275.375.0
PIDNet-L74.074.975.374.6

Table S1. Sensitivity to the auxiliary-loss weight λaux\lambda_{\mathrm{aux}} across semantic segmentation models (Foggy Cityscapes val, mean fog mIoU, %); the λaux=0.4\lambda_{\mathrm{aux}} = 0.4 default consistently achieves optimal performance (bold).

Design Ablations

Extensive ablations on the Foggy Cityscapes val set validate our structural and hyperparameter choices against unguided baselines, which achieve 58.0–63.4% mean mIoU across the four base models. The full LiDSeg configuration improves the BiSeNetV2 baseline by 17.4 mIoU points on Foggy Cityscapes and 17.6 points on ACDC pooled-All validation. Reconstruction supervision and LGF provide complementary gains: reconstruction with concatenation improves the baseline by 8.1/9.0 points on Foggy Cityscapes/ACDC, while LGF without reconstruction improves it by 8.68/11.0 points. Combining reconstruction supervision with LGF achieves 75.4%/73.6% mIoU. Among fusion alternatives, spatial and cross-attention reach 75.6% and 75.9% mIoU, respectively, compared with 75.4% for LGF, but require higher parameter and computational cost, giving LGF a favorable accuracy-efficiency trade-off for edge deployment.

The asymmetric design further improves efficiency through linear attention, which matches Softmax performance at approximately 4×4 \times fewer FLOPs. Multiscale guidance is also important: fusing all three feature strides (s8, s16, and s32) consistently outperforms the tested single- and two-scale configurations (Table S2). The default (3,3,2) guidance-encoder capacity provides a strong efficiency–accuracy balance; increasing it to (4,4,3) adds 2.4M parameters for only a 0.5 mIoU-point gain (Table S2). Training choices also affect robustness (Table S2). Dynamic uncertainty weighting improves mean mIoU by 0.8 points across all four models compared with equal weighting, while the progressive fog curriculum improves performance by 2.3 mIoU points over joint training across all fog densities.

SettingBiSeNetV2DDRNet-23sSTDC1-SegPIDNet-L
(a) Multi-scale fusion
none (baseline)58.061.562.063.4
s869.570.670.871.3
s1670.471.371.571.9
s3267.869.369.570.1
s8+s1673.673.973.974.1
s16+s3273.173.573.573.7
s8+s16+s32 (default)75.475.375.375.3
(b) Guidance-encoder capacity
Tiny (2,2,1)72.872.772.772.7
Default (3,3,2) (default)75.475.375.375.3
Base (4,4,3)75.975.875.875.8
(c) Loss weighting
equal weights74.674.574.574.5
uncertainty (default)75.475.375.375.3
(d) Fog curriculum
joint (all densities)73.173.073.073.0
progressive (default)75.475.375.375.3

Table S2. Design ablations across all four real-time models (Foggy Cityscapes val, mIoU % over the three fog densities). Each block varies one design choice; Fusion “none” is the unguided baseline.

Additional Experimental Result Analysis

Foggy Cityscapes: Per Class Performance

We extend the analysis of the Foggy Cityscapes results to demonstrate LiDSeg’s per-class robustness (Table S3). Semantic preservation directly correlates to the shared latent structure learned via the auxiliary reconstruction task, as PSNR and SSIM tightly co-vary with segmentation mIoU throughout training (Table S4). The gains achieved by LGF are largely driven by the physical nature of these weather effects. By measuring the cross-channel Pearson correlation (ρˉ\bar{\rho}) of the clear-to-degraded residuals, we observe that fog (ρˉ=0.86\bar{\rho} = 0.86), rain (ρˉ=0.83\bar{\rho} = 0.83), and snow (ρˉ=0.81\bar{\rho} = 0.81) degrade images coherently across channels, making channel-based suppression highly effective. However, nighttime scenes exhibit lower coherence (ρˉ≈0.50\bar{\rho} \approx0.50) due to non-uniform artificial lighting, explaining why night remains fundamentally the most challenging condition to modulate via channel-wise gating.

MethodRoadSide.Buil.WallFenc.PoleLigh.SignVege.Terr.SkyPers.Ride.CarTruc.BusTrai.Moto.Bicy.mIoU
LiDSeg (BiSeNetV2)9883905855626674926695766093667670567474.2
LiDSeg (DDRNet-23s)9982905754626673926795756093697571557374.1
LiDSeg (STDC1-Seg)9884895855626473926795756194687471577574.3
LiDSeg (PIDNet-L)9984895753626674926696785894677570577374.2
LiDSeg (TopFormer-B)9782895754626675926595745992667669547473.6
LiDSeg (SeaFormer-B)9782895653616674916695735992657670547473.3

Table S3. Per-class IoU (%) on Foggy Cityscapes val at dense fog (β=0.02\beta= 0.02) for all LiDSeg variants, including the transformer extensions. Bold: best per class.

β\betaPSNRSSIMmIoU
0.00534.20.94676.3
0.0133.30.92975.7
0.0231.00.90274.2

Table S4. Reconstruction quality vs. fog density (BiSeNetV2). Input PSNR is foggy vs. clear.

Because sequential restoration optimizes for pixel fidelity rather than semantic discriminability, it can accumulate restoration errors, which LiDSeg avoids propagating by removing explicit restoration at inference. This efficient, guidance-driven approach also achieved zero-shot generalizability where LiDSeg variants achieve a consistent +8 to +15 mIoU-point improvement over baselines on both the Foggy Zurich v2 and Foggy Driving datasets (Table S6).

Semantic segmentation modelFZ (unguided)FZ (LiDSeg)FD (unguided)FD (LiDSeg)
BiSeNetV238.253.541.050.4
DDRNet-23s40.152.842.551.0
PIDNet-L39.552.641.850.8
STDC1-Seg39.053.442.052.4

Table S6. Zero-shot mIoU (%) on Foggy Zurich v2 (FZ) and Foggy Driving (FD): unguided Semantic segmentation model vs. LiDSeg (no target labels).

Beyond synthetic fog analysis, LiDSeg proves to be a robust alternative to fundamental paradigm designs, fully capable of real-world deployment. For instance, pipelining an explicit dehazing model before segmentation (the restore-then-segment paradigm) results in heavy latency and actually underperforms on mIoU (Table S5).

PipelineParamsGFLOPsFPSmIoU
DehazeFormer →\rightarrow BiSeNetV25.9M61.45471.5
Restormer →\rightarrow BiSeNetV229.4M213.01872.3
LiDSeg (BiSeNetV2)5.22M27.283.474.2

Table S5. LiDSeg vs. restore-then-segment with the same BiSeNetV2 head. GFLOPs/FPS at 512×1024512 \times1024 (Jetson AGX Orin); mIoU at dense fog (β=0.02\beta= 0.02).

Most importantly, these robustness improvements do not compromise edge deployability. Complementing the 512×1024 metrics provided in the main paper, lower-resolution (256×512) evaluations confirm that LiDSeg comfortably exceeds real-time thresholds (138–587 FPS) on a Jetson AGX Orin (Table S7).

ModelVariantParams (M)GFLOPsFPSMem
BiSeNetV2Baseline3.366.23161.518
DDRNetBaseline5.702.29707.318
STDCBaseline8.4310.39724.223
SCTNet-BBaseline17.447.04509.344
PIDNet-LBaseline37.3117.26308.080
TopFormer-BBaseline7.062.42715.224
SeaFormer-BBaseline8.580.89252.227
LiDSeg (BiSeNetV2)Guided5.226.80138.221
LiDSeg (DDRNet)Guided7.652.88587.123
LiDSeg (STDC)Guided11.4811.02545.133
LiDSeg (PIDNet-L)Guided39.5417.87249.488
LiDSeg (TopFormer-B)Guided9.013.01512.828

Table S7. Jetson AGX Orin efficiency at 256×512256 \times512 (Benchmarking more efficiency in addition to 512×1024512 \times1024 Table 6). Guided: LiDSeg with the reconstruction decoder discarded at inference.

ACDC: Expert Cross-Condition Generalization

To isolate the value of joint training and expose the channel-coherent structure of weather degradation, Table S8 reports a full cross-condition evaluation of the four condition-specific Expert LiDSeg models. Each Expert was fine-tuned on a single ACDC split and evaluated on all four. The 4×4 matrix per guided model shows a pattern consistent with our channel-correlation analysis. Feature transfer within the channel-coherent F↔R↔S triangle is notably strong. Because fog, rain, and snow share a structurally coherent channel degradation characteristic (ρˉ>0.8\bar{\rho} > 0.8), any Expert trained on one of these splits retains a large fraction of its diagonal accuracy on the other two. The best off-diagonal entry per model consistently lies inside this triangle, confirming that a shared encoder generalizes well across weather corruptions.

Similar to previous results, night transfer is bounded by scene priors rather than condition-specific features. Every Fog/Rain/Snow Expert evaluated on Night drops below its own diagonal into a narrow 36–43% range. However, this is still substantially above what a Cityscapes-only Source model achieves on ACDC-Night (∼12\sim12–14%) and comparable to the zero-shot LiDSeg performance on Foggy-Zurich and Foggy-Driving. The symmetric N→\rightarrowF/R/S rows convey the same point from the opposite direction: Expert Night transfers acceptably to Fog (46.5–50.4%) but degrade on Rain and Snow (39.3–43.4%), because nighttime supervision under-samples the high-illumination feature space that those conditions depend on. Overall, the Uber joint fine-tuning reported in the main paper demonstrates that pooling channel-coherent supervision across all conditions strengthens the shared encoder simultaneously, learning generalized features that natively benefit even disjointed domains

like Night.

Qualitative Results

Finally, visual comparisons ground our quantitative findings. Figure S1 and Figure S2 provide qualitative evidence on both synthetic Foggy Cityscapes and real-world ACDC. Foggy Cityscapes includes the training-only reconstruction output alongside segmentation, while ACDC shows segmentation outputs only because no reconstruction decoder is used during ACDC fine-tuning or inference. The examples illustrate preservation of details such as lane markings, pedestrians, and distant objects under adverse conditions.

Foggy Cityscapes reconstruction and segmentation comparison

Figure S1. Foggy Cityscapes val: reconstruction and segmentation. Four Semantic segmentation models at medium (β=0.01\beta= 0.01) and dense (β=0.02\beta= 0.02) fog. Columns: foggy input, LiDSeg reconstruction (training only), clear reference, ground truth, unguided baseline, LiDSeg.

Real ACDC val segmentation results for all conditions and semantic segmentation models

Figure S2. Real ACDC val: segmentation, all conditions for all Semantic segmentation models. Sixteen distinct scenes (Fog, Night, Rain, Snow and four Semantic segmentation models). Columns: adverse input, ground truth, unguided baseline, LiDSeg.

References

  1. [1]Maxim Berman et al. The Lovász-Softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In Proc. CVPR, pages 4413–4421, 2018.
  2. [2]O. Bi et al. Generalized foggy-scene semantic segmentation by frequency decoupling. In Proc. CVPR Workshops, 2024.DOI
  3. [3]L. Chen et al. RT-LinearFormer: A real-time linear vision transformer for semantic segmentation. Neurocomputing, 2025.
  4. [4]Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proc. ECCV, 2018.arxiv.org/abs/1802.02611
  5. [5]Zhe Chen, Yuche Duan, Yu-Gang Wen, Junjoseph, Kui Jia, and Jian Gao. ViT-Adapter: Training plain vision transformers for dense prediction. In Proc. NeurIPS, 2022.
  6. [6]Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Garg. Masked-attention mask transformer for universal image segmentation. In Proc. CVPR, 2022.arxiv.org/abs/2112.01527
  7. [7]Sungha Choi, Sanghun Jung, Hunsang Yun, Joanne T. Kim, Seungryong Kim, and Jaegul Choo. RobustNet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In Proc. CVPR, 2021.arxiv.org/abs/2103.15597
  8. [8]G. Csurka et al. Semantic image segmentation: Two decades of research. Found. Trends Comput. Graph. Vis., 2023.
  9. [9]Yann N. Dauphin, Angela Fan, Michael Auli, and David Grangier. Language modeling with gated convolutional networks. In Proc. ICML, pages 933–941, 2017.
  10. [10]Mingyuan Fan, Shengqi Lai, Junshi Huang, Xiaoming Wei, Zhenhua Chai, Junfeng Luo, and Xiaoming Wei. Rethinking BiSeNet for real-time semantic segmentation. In Proc. CVPR, 2021.arxiv.org/abs/2104.13188
  11. [11]Y. Ganin et al. Domain-adversarial training of neural networks. JMLR, 17(59):1–35, 2016.arxiv.org/abs/1505.07818
  12. [12]K. Garg and S.K. Nayar. Vision and rain. IJCV, 58(2):135–148, 2004.
  13. [13]Minseok Han, Jungwuk Kim, Kyungmin Lee, and Jaehyun Park. HALO: Hyperbolic active learning for semantic segmentation under domain shift. In Proc. ICML, 2024. Adverse conditions segmentation track.
  14. [14]Yuanduo Hong, Huihui Pan, Weichao Sun, and Yisong Jia. Deep dual-resolution networks for real-time and accurate semantic segmentation of road scenes. In Proc. CVPR, 2021.arxiv.org/abs/2101.06085
  15. [15]Lukas Hoyer, Dengxin Dai, and Luc Van Gool. DAFormer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In Proc. CVPR, 2022.arxiv.org/abs/2111.14887
  16. [16]Lukas Hoyer, Dengxin Dai, and Luc Van Gool. HRDA: Context-aware high-resolution domain-adaptive semantic segmentation. In Proc. ECCV, 2022.arxiv.org/abs/2204.13132
  17. [17]Lukas Hoyer, Dengxin Dai, Haoran Wang, and Luc Van Gool. MIC: Masked image consistency for context-enhanced domain adaptation. In Proc. CVPR, 2023.arxiv.org/abs/2212.01322
  18. [18]Jie Hu, Li Shen, Samuel Albanie, Gang Sun, and Enhua Wu. Squeeze-and-excitation networks. IEEE TPAMI, 42(8):2011–2023, 2020.
  19. [19]J. Huang et al. SPC: Style-preserving contrastive learning for domain generalized semantic segmentation. In Proc. CVPR, 2023.
  20. [20]Lei Huang, Yi Zhou, Fan Zhu, Li Liu, and Ling Shao. Iterative normalization: Beyond standardization towards efficient whitening. In Proc. CVPR, 2019.arxiv.org/abs/1904.03441
  21. [21]Angelos Katharopoulos et al. Transformers are RNNs: Fast autoregressive transformers with linear attention. In Proc. ICML, pages 5156–5165, 2020.arxiv.org/abs/2006.16236
  22. [22]Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proc. CVPR, 2018.arxiv.org/abs/1705.07115
  23. [23]Wei Kou et al. Adverse weather-independent framework towards autonomous driving perception through temporal correlation and unfolded regularization. arXiv preprint arXiv:2508.01583, 2025.arxiv.org/abs/2508.01583
  24. [24]Chaehyeon Lee, Jae Young Cho, Jaehoon Kim, Dong-Geol Kim, and Jaegul Choo. FIFO: Learning fog-invariant features for foggy scene segmentation. In Proc. CVPR, 2022.arxiv.org/abs/2204.01587
  25. [25]Boyi Li, Wेंqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhiyong Wang. Benchmarking single-image dehazing and beyond. IEEE TIP, 28(1):492–505, 2019.DOI
  26. [27]Guosheng Lin, Anton Milan, Chunhua Shen, and Ian Reid. RefineNet: Multi-path refinement networks for high-resolution semantic segmentation. In Proc. CVPR, 2017.arxiv.org/abs/1611.06612
  27. [28]Yang Liu, Haoran Zhang, Wenyu Li, and Wanli Liu. CoDA: Chain-of-domain adaptation for adverse weather semantic segmentation. In Proc. ECCV, 2024.
  28. [29]X. Ma et al. Both style and fog matter: Cumulative domain adaptation for semantic foggy scene understanding. In Proc. CVPR, 2024.
  29. [30]Srinivasa G. Narasimhan and Shree K. Nayar. Vision and the atmosphere. IJCV, 48(3): 233–254, 2002.
  30. [31]Shree K. Nayar and Srinivasa G. Narasimhan. Vision in bad weather. In Proc. ICCV, volume 2, pages 820–827, 1999.DOI
  31. [32]Vladimir Nekrasov, Thanuja Dharmasiri, Andrew Spek, Tom Drummond, Chunhua Shen, and Ian Reid. Light-weight RefineNet for real-time semantic segmentation. In Proc. BMVC, 2018.arxiv.org/abs/1810.03272
  32. [33]X. Niu et al. Effective image restoration for semantic segmentation. Neurocomputing, 2020.DOI
  33. [34]X. Pan et al. Two at once: Enhancing learning and generalization capacities via IBN-Net. In Proc. ECCV, 2018.arxiv.org/abs/1807.09441
  34. [35]Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaoou Tang, and Ping Luo. Switchable whitening for deep representation learning. In Proc. CVPR, 2019.arxiv.org/abs/1904.09739
  35. [36]Y. Pan et al. Exploring weather-aware aggregation and adaptation for semantic segmentation under adverse conditions. In Proc. ICCV, 2025.DOI
  36. [37]D Pushpalatha and P Prithvi. LDFormer: lightweight dehazing transformer. Neural Computing and Applications, pages 1–20, 2025.DOI
  37. [38]Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Map-guided curriculum domain adaptation and uncertainty-aware evaluation for semantic nighttime image segmentation. IEEE TPAMI, 2020.arxiv.org/abs/2005.14553
  38. [39]Christos Sakaridis, Dengxin Dai, and Luc Van Gool. ACDC: The adverse conditions dataset with correspondences for semantic driving scene understanding. In Proc. ICCV, 2021.DOI
  39. [40]Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Curriculum model adaptation with synthetic and real data for semantic foggy scene understanding. IJCV, 2021.
  40. [41]Christos Sakaridis et al. Semantic foggy scene understanding with synthetic data. IJCV, 126(9):973–992, 2018.DOI
  41. [42]David R. So, Wojciech Mańke, Hanxiao Liu, Zihang Dai, Noam Shazeer, and Quoc V. Le. Primer: Searching for efficient transformers for language modeling. In Proc. NeurIPS, pages 11881–11895, 2021.arxiv.org/abs/2109.08668
  42. [43]Y. Song et al. Vision transformers for single image dehazing. TIP, 2023.DOI
  43. [44]Xudong Sun et al. D2SL: Decouple defogging and semantic learning for foggy domain-adaptive semantic segmentation. arXiv preprint arXiv:2404.04807, 2024.
  44. [45]Yi-Hsuan Tsai et al. Learning to adapt to structured output space for semantic segmentation. In Proc. CVPR, pages 7472–7481, 2018.arxiv.org/abs/1802.10349
  45. [46]Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M. Patel. TransWeather: Transformer-based restoration of images degraded by adverse weather conditions. In Proc. CVPR, 2022.arxiv.org/abs/2111.14813
  46. [47]Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. ADVENT: Adversarial entropy minimization for domain adaptation in semantic segmentation. In Proc. CVPR, 2019.arxiv.org/abs/1811.12833
  47. [48]Qiang Wang, Zilong Huang, Jiangwei Lu, Gang Yu, and Li Zhang. SeaFormer: Squeeze-enhanced axial transformer for mobile semantic segmentation. In Proc. ICLR, 2023.
  48. [49]J. Wang et al. A two-stage adverse weather semantic segmentation method for weatherproof challenge CVPR 2024 workshop UG2+. arXiv preprint arXiv:2406.05513, 2024.arxiv.org/abs/2406.05513
  49. [50]Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yuning Ding, Zhanpeng Guo, Han Xu, et al. Deep high-resolution representation learning for visual recognition. IEEE TPAMI, 43(10):3349–3364, 2020.DOI
  50. [51]Zhizheng Wu, Xiao Han, Yuxin Lin, Mustafa Gokhan Uzunbas, Tom Goldstein, Ser Nam Lim, and Dimitris N. Metaxas. DANNNet: A one-stage domain adaptation network for unsupervised nighttime semantic segmentation. In Proc. CVPR, 2021.arxiv.org/abs/2104.10834
  51. [52]Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. SegFormer: Simple and efficient design for semantic segmentation with transformers. In Proc. NeurIPS, 2021.arxiv.org/abs/2105.15203
  52. [53]Jiacong Xu, Zixiang Xiong, and Shankar P. Bhatt. PIDNet: A real-time semantic segmentation network inspired by PID controllers. In Proc. CVPR, 2023.arxiv.org/abs/2206.02066
  53. [54]X. Yang et al. Semantic segmentation in multiple adverse weather conditions with domain knowledge retention. In Proc. AAAI, 2024.DOI
  54. [55]Yanchao Yang and Stefano Soatto. FDA: Fourier domain adaptation for semantic segmentation. In Proc. CVPR, 2020.
  55. [56]Changqian Yu, Changxin Gao, Jingbo Wang, Gang Yu, Chunhua Shen, and Nong Sang. BiSeNet V2: Bilateral network with guided aggregation for real-time semantic segmentation. IJCV, 129(11):3051–3068, 2021.arxiv.org/abs/2004.02147
  56. [57]S.W. Zamir et al. Restormer: Efficient transformer for high-resolution image restoration. In Proc. CVPR, 2022.arxiv.org/abs/2111.09881
  57. [58]Wenqiang Zhang, Zilong Huang, Guozhong Luo, Tao Chen, Xingang Wang, Wenyu Liu, Gang Yu, and Chunhua Shen. TopFormer: Token pyramid transformer for mobile semantic segmentation. In Proc. CVPR, 2022.arxiv.org/abs/2204.05525
  58. [59]Yurui Zhang, Gaofeng Huang, Jie Cao, Jiwen Luo, Jingshan Xue, and Chunlei Zhou. WeatherStream: Light transport automation of single image deweathering. In Proc. CVPR, 2023.DOI
  59. [60]Hengshuang Zhao, Jianping Shi, Xiaojun Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proc. CVPR, pages 2881–2890, 2017.

Paper details

vlnc:W_tb5wfdg9
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents