betaFeedback

WorksComputer Science

Passable for Whom? Few-Shot Personal Calibration of Sidewalk Accessibility Perception

Abstract

Whether a sidewalk is passable depends on who is trying to pass: people who use the same mobility aid disagree about the same image, so one label per scene, or per aid, is wrong for many of them. We introduce few-shot personal accessibility perception: from K answers of a person never seen before, predict their judgment of sidewalk scenes never seen before. PassCal separates a shared description of the scene from the person's interpretation of it. The description comes from the internal state of a frozen vision-language model (VLM) and from barrier type, severity and context properties predicted by a detector trained on 14.9k public audits; the interpretation, a calibration of the person's response style and a weighting of the properties, is fit from their answers. On 8,883 judgments in which each of 135 verified mobility-aid users rated the same 50 sidewalk scenes, with people and scenes both held out, eight answers lower the Brier score from 0.550 to 0.479. The same VLM prompted with those answers reaches 0.562, and the same calibration without the scene 0.521. Conformal risk control turns predictions into passable, blocked or "depends on you" statements with a bounded expected rate of wrong statements over people; at 5%, eight answers raise the share of decisive scenes from 27% to 42%. Audit-supervised properties improve all sixteen vision and VLM encoders we test and outperform attributes from a sidewalk navigation VLM.

Introduction

A missing curb ramp stops one wheelchair user from egress and is a short detour for another. A cracked slab that a cane user steps over is a hazard for someone using a walker. Accessibility maps and routing tools [6, 19, 40, 44] never theless attach one label to each place, and learned perception systems [21, 27, 30] are trained to reproduce that label. Li et al. [24] showed images of sidewalk barriers to people who use canes, walkers, mobility scooters, and manual and motorized wheelchairs, and found large differences between these groups. In their data we find that differences within a group are as large (Figure 1): for cane and walker users, a person’s own tendency explains more of the variance in their answers than the image does (27% against 14% and 18%; supplement).

Figure 1. The same sidewalk, different answers. Three survey scenes with the answers of cane users and manual-wheelchair users, and the properties recorded by Project Sidewalk auditors (audited) and by our detector (predicted). People disagree between aids and within an aid. PASSCAL describes the scene and learns from a few answers what that description means for one person.

Judgments like these are hard to collect. In the case of [24], each one comes from a person who actually uses a cane, walker, wheelchair or scooter, reached through disability organisations over two months, paid for a 25-minute survey, and screened against the automated and incentive-driven sign-ups that paid disability studies attract: Li et al. removed 5,239 fraudulent sign-ups and 68 suspicious completed surveys to obtain theirs [24]. The resulting 8,883 answers represent about 56 hours of mobility-aid users’ time, from more people than in most accessibility studies (the median at CHI and ASSETS is 13 [28]), and the design is dense: every person rates the same 50 scenes, which separates what a person brings from what the scene shows. It is, to our knowledge, the only public dataset of individual sidewalk passability judgments by mobility-aid users. Every new scene must be judged by such people, so the cost lies in the answers, not in the images: a method for this domain should be judged by how much it extracts from a few real answers per person.

This paper asks what a vision system should do when population labels miss this much individual variation and the judgments needed to fix it are this costly to collect. We separate a shared description of the scene from a person’s interpretation of it. Public audits and a frozen vision-language model (VLM) supply the description; a short interaction supplies the personal answers that an aid label cannot determine. The target is the judgment a person reports about a photograph. The formulation has three parts.

  1. Perception. A detector trained on public Project Sidewalk audits [39] predicts the barrier type, its severity, and context tags for a street-view image, and a frozen VLM asked whether the sidewalk is passable provides a scene embedding. No user judgment is involved.

  1. Personal calibration. A person answers KK questions about calibration images just once. From these answers we estimate personal weights on the properties and a calibration of the person’s response style.

  1. Statements with controlled risk. Instead of a probability, the system says “passable”, “blocked”, or “depends on you”. Conformal risk control [1] bounds how often each kind of statement is wrong, in expectation over people.

We evaluate with both the person and the query images excluded from model fitting, which tests whether answers on calibration images improve predictions on other scenes, avoiding overfitting to a familiar rater or a scene’s crowd label. Our method, PassCAL, learns how a person departs from a model of everyone else, with three response parameters and ten property weights. The two obvious baseline alternatives do not: despite being given eight user answers in its prompt, a recalibrated VLM is no better than a model that ignores the person, although the same frozen VLM used as a scene encoder inside PassCAL gives our best result; and a regressor fit per person on the VLM state, as done for personal aesthetics [38], cannot learn a person from eight answers.

Our contributions are:

  • A benchmark for few-shot personal accessibility perception: 8,883 dense judgments from 135 mobility-aid users, a protocol that holds out people and scenes, intervals that resample both, a held-out barrier type, and baselines from in-context VLMs, personal aesthetics, crowd oracles and a pedestrian navigation VLM. We will release code and splits.

  • PassCAL, which separates an audit-supervised scene description from a few-shot personal interpretation. Eight answers reduce the Brier score by 0.070, and PassCAL outperforms in-context personalisation of the same VLM by 0.090 and per-person regression on its state by 0.068.

  • Evidence that audit supervision adds what foundation models miss: predicted sidewalk properties improve all sixteen vision and VLM encoders we test, and an oracle study locates the remaining gains in barrier type and severity.

  • Statements whose expected rate of wrong answers per scene is bounded over people; at a 5% bound, eight answers raise the decisive share from 27% to 42%.

Related Work

Sidewalk accessibility from street-level imagery. Crowd-sourced audits of Street View [17, 18, 39] have been scaled with classifiers and detectors [27, 30, 47] and complemented by aerial extraction [20, 21] and multimodal sensing [9]. Recent studies test vision-language models on accessibility attributes: they estimate sidewalk width reasonably but slope and pavement poorly [25], and miss subtle surface conditions [46], and grounded VLMs now converse about pedestrian navigation [43]. These works predict a property of the place, but we predict the judgment of a person.

Personalised accessibility. Routing systems accept hand-set user weights [6, 40, 44]. Li et al. [24] collected the judgments we use and analysed differences between mobility groups statistically, without a predictive model of individuals. Accessibility Scout [22] personalises LLM-generated scans of photos and evaluates them on 500 images and in a 10-person user study, without a predictive model of individual judgments. We contribute a quantitative protocol and a model at the individual level.

Personalised and subjective visual prediction. Personalised image aesthetics adapts a generic model with a few ratings per user [37, 50, 51, 53]; Ryu and Yanaka [38] fit per-user linear heads on the hidden states of a frozen VLM, the closest analogue to our use of a VLM state. Few-shot personalisation also predicts an unseen person’s gaze [48, 49], and a few answers from a new user reweight shared attributes or earlier users’ preferences in generation and interface design [3, 33]. We differ in three ways: the shared description is supervised by audits of the property in question, the person’s response style is calibrated separately from their weights, and the output is a statement with a bounded error rate, evaluated on people and images that are both new. Individual quality models learn observer-specific responses to distortions [14], and individual preferences are represented unevenly [15]. Urban perception has moved from population scores [13] to differences across demographic groups [35] and to models conditioned on rater attributes [10]. We condition on a person’s answers instead of their attributes, in a domain where an error has a physical cost. Work on learning from disagreement treats annotator variation as signal [8, 11, 34]; ours is the few-shot, unseen-annotator case.

Concept-based models. Concept bottleneck models predict through human-readable concepts [23], with a residual on the embedding [52] or interactive correction [7, 42]. Our concepts come from an external audit corpus, and what is personal is the weight a user gives them.

Calibration and conformal prediction. Temperature scaling corrects the confidence of a trained classifier [16]; we fit such a correction per person from a few answers. Conformal risk control bounds the expectation of a monotone loss with exchangeable calibration data [1], few-shot variants study per-task guarantees [32], conformal intervals have been attached to VLM estimates of sidewalk attributes [25], and group-conditional versions control risk within each group of a hierarchy with abstention [41]. We apply it with people as the exchangeable units.

Data

Judgments. We use the public release of Li et al. [24]. Respondents saw street-view images of sidewalk barriers taken from Project Sidewalk and answered, for the mobility aid they use, whether they could pass: yes, unsure, or no. After removing 5,239 fraudulent sign-ups and 68 suspicious completed surveys, the release keeps 190 survey responses. Of these, 178 person-aid profiles from 135 people concern one of the five aids (some people use more than one): walking cane 53, manual wheelchair 37, motorized wheelchair 34, walker 32, mobility scooter 22. We use the 50 images that we could match to their Project Sidewalk record, with 8,883 judgments (50% yes, 38% no, 12% unsure).

Objective properties. Each survey image is a Project Sidewalk label, which carries a label type (curb ramp, missing curb ramp, obstacle, surface problem), a severity from 1 to 3, and free tags. We group tags into five (narrow, vegetation, surface damage, object, no alternate route), which gives a 10-dimensional property vector. We call the values recorded by auditors audited properties and the detector’s outputs predicted properties.

How people differ. In only 36 of the 260 image-aid cells do at least 80% of respondents agree. No property is a universal dealbreaker, and the shared core is directional: higher severity lowers passability for every aid, while for surface problems, narrow sidewalks and objects 30 to 60% of people have a personal effect opposite to that of their group.

Audit pool. To train the property detector we sampled Project Sidewalk labels from nine cities that include the eight cities of the survey images through its public API, stratified by city, type and severity. We excluded every survey label and every label on a panorama that contains one, and kept labels whose panorama still resolves in the Street View Static API. The pool has 14,893 scenes; an extension to 18,092 is used in the scaling analysis and in § 5.8.

Method

Property detector

A Project Sidewalk label is a point in a wide scene. From the label’s position on the panorama we compute its pixel in the downloaded 90° view (median error 0.1 px against the auditor’s recorded click, on 6,000 labels) and take a 50° crop centred on it. A second crop, from the half of the image away from the label, serves as a background example. We fine-tune the last four blocks of DINOv2-L [31] with three heads: five-way type (four label types plus background), three-way severity on label crops, and five binary tags. Validation is grouped by panorama.

At test time the label location is not used, since a deployed system would not have it. The image is covered by square windows at three scales. Each window is scored, and per-image properties are the mean of the three highest-scoring windows per output, with severity weighted by the probability that a window is not background. We average three seeds.

Scene embedding

The properties name what auditors record, not everything that bears on passability. For the rest we use a frozen VLM, Qwen3-VL-8B [4], as an encoder. The model receives the image and a question asking whether a person who uses a mobility aid could pass through this sidewalk (full prompt in the supplement), and we keep the last-layer hidden state at the position where it would begin its answer, discarding the answer itself. A 16-dimensional PCA of this state, fit on 3,000 pool scenes, is the scene embedding ziz_i. The VLM never sees a user’s answers and is not fine-tuned. Section 5.5 compares this choice with fifteen other encoders.

PASSCAL system diagram

Figure 2. PassCal. The detector and the frozen VLM describe the scene without any user judgment. A few answers from the person set the personal parameters, and thresholds calibrated on other people turn the prediction into a statement with bounded risk.

Personal model

Let xi=(zi,si)∈R26x_i=(z_i,s_i)\in\mathbb{R}^{26} represent image ii: the scene embedding concatenated with the 10 predicted properties sis_i. For a person pp using aid aa, the answer follows a proportional-odds model [29] with latent passability

ηp,i=waTxi+δpTsi(1)\eta_{p,i}=w_a^{\mathsf{T}}x_i+\delta_p^{\mathsf{T}}s_i \tag*{(1)}

and aid-specific thresholds τa,0<τa,1\tau_{a,0}<\tau_{a,1}, so that Pp,i(y≤c)=σ(τa,c−ηp,i)P_{p,i}(y\le c)=\sigma(\tau_{a,c}-\eta_{p,i}) for c∈{no,unsure}c\in\{\text{no},\text{unsure}\}. The group weights waw_a and the thresholds are fit on training people. The group term uses the whole representation; the personal weights δp\delta_p act only on the properties, so that what the system learns about how a person sees a scene is expressed over named quantities.

People also differ in how they respond, independently of the scene: some say no more readily, some rarely say unsure, some are more consistent than the group. We model this with a personal response calibration of the predicted distribution,

qp,i(c)∝exp⁡(βplog⁡Pp,i(c)+op,c),(2)q_{p,i}(c)\propto\exp\left(\beta_p\log P_{p,i}(c)+o_{p,c}\right), \tag*{(2)}

with class offsets opo_p that sum to zero and a positive inverse temperature βp\beta_p, a per-person form of temperature scaling [16]. The offsets capture strictness and the use of “unsure”; βp\beta_p captures confidence.

Both sets of personal parameters are estimated by MAP from the person’s KK calibration answers, first the property weights,

min⁡δp−∑k=1Klog⁡Pp,k(yp,k)+λδ∥δp∥2,(3)\min_{\delta_p}-\sum_{k=1}^{K}\log P_{p,k}(y_{p,k})+\lambda_\delta\lVert\delta_p\rVert^2, \tag*{(3)}

and then (op,βp)(o_p,\beta_p) on the same answers with Gaussian priors centred on op=0o_p=0 and βp=1\beta_p=1. With K=0K=0 the model returns the group prediction. The prior strengths are tuned on training people only. Section 5.4 compares this choice with simpler personal parameterisations.

Statements with controlled risk

The system says PASSABLE if q(yes)≥t+q(\text{yes})\ge t_+, BLOCKED if q(no)≥t−q(\text{no})\ge t_-, and DEPENDS otherwise. The harmful error for PASSABLE is saying it to a person who answers no. For a person pp with held-out scenes SpS_p, define

Lp(t)=1∣Sp∣∑i∈Sp1{qp,i(yes)≥t}1{yp,i=no},(4)L_p(t)=\frac{1}{\lvert S_p\rvert}\sum_{i\in S_p}\mathbf{1}\{q_{p,i}(\text{yes})\ge t\}\mathbf{1}\{y_{p,i}=\text{no}\}, \tag*{(4)}

which is non-increasing in tt. Given nn calibration people who went through the same KK-answer procedure, conformal risk control [1] selects the smallest t+t_+ with nn+1L‾(t+)+1n+1≤α\frac{n}{n+1}\overline{L}(t_+)+\frac{1}{n+1}\le\alpha, which guarantees E[Lnew(t+)]≤α\mathbb{E}[L_{\mathrm{new}}(t_+)]\le\alpha for a new exchangeable person. The threshold t−t_- is set symmetrically for saying BLOCKED to a person who answers yes; if both conditions hold, PASSABLE takes precedence, which can only remove BLOCKED statements, so its bound still holds. An answer of unsure counts as neither error. The guarantee is marginal over people, not conditional on a person or an aid. Each statement type is bounded separately at α\alpha, and the denominator is all of a person’s scenes, not only the scenes that receive that statement. And calibration people are scored on the same held-out images as test people, so the bound covers new people on a fixed set of scenes. We report the decisive share, the fraction of scenes that receive PASSABLE or BLOCKED, at fixed α\alpha.

Experiments

Protocol

People and images are each split into five folds. In fold ff the group model is fit on training people and training images. A test person gives KK answers on training images, as with a fixed calibration set shipped with an application, and is evaluated on the held-out images, whose votes were excluded from group-model fitting. The metric is the Brier score of the predicted distribution against the person’s answer (summed over the three classes), averaged within a person and then over people. We repeat the whole procedure for 20 random partitions of people and images, with three draws of calibration items each, and report means over partitions; the standard deviation of a table entry across partitions is 0.007 to 0.010. Comparisons are paired by person, using each person’s mean over partitions: bootstrap 95% intervals over the 135 people and Wilcoxon signed-rank tests. Because all partitions share the same 50 images, we also report a two-way bootstrap that resamples people and images together (supplement), and a split that withholds a whole barrier type (Sec. 5.6). For statements, the non-test people are split in half into people who fit the model and people who calibrate thresholds, repeated 20 times. Unless stated otherwise, every number in this section comes from this protocol.

References. Embedding only uses the scene embedding alone, with personal weights on its dimensions. Audited properties replaces the predicted properties with the auditors’. Crowd oracle predicts the distribution of the 18 to 42 training people’s answers on the test image for the same aid, which a deployed system would not have; collaborative filtering needs the same votes on the query scene. Crowd oracle + strictness adds a personal strictness to that oracle, a reference no deployable system can reach. Per-person regression is the Linear-Hidden model of Ryu and Yanaka [38]: a ridge regressor on the VLM state fit on the person’s own answers only. No image fits the same response calibration to the aid’s answer distribution, without the scene. WalkGPT attributes replaces our properties with ten attributes parsed from the answers of WalkGPT-13B [43] on each image (supplement). In-context VLM gives the same Qwen3-VL-8B the person’s eight calibration images with their answers and asks about the test image; we read the next-token distribution over the three answers.

Main results

Personalisation works on unseen sidewalks. Table 1 and Figure 3 give the main result. Eight answers reduce the Brier score of PassCAL by 0.070 (CI [0.051, 0.092]), better for 68% of people) and twenty by 0.093 (79%), in all 20 partitions; resampling images as well as people leaves both intervals clear of zero (supplement).

ModelK=0K=0K=3K=3K=8K=8K=20K=20
Per-person regression [38]–0.5660.5500.521
No image + response cal.0.5830.5480.5210.503
WalkGPT [43] attributes0.5560.5190.4900.473
Embedding only0.5520.5140.4830.464
PassCAL (ours)0.5500.5120.4790.457
Audited properties0.5370.4970.4630.438
Crowd oracle0.514
Crowd oracle + strictness0.5140.4740.4500.437

Table 1. New people on new sidewalks: 135 people, 50 images, mean over 20 partitions. Per-person regression uses the first calibration draw (PassCAL scores 0.514, 0.480 and 0.458 on the same items). Rows below the line use information that a deployed system does not have.

Figure 3. Brier score against the number of calibration answers (mean over 20 partitions). The diamond is the same VLM with the eight answers in its prompt, after recalibration (five partitions).

The mean score is below the crowd oracle’s by 0.034 at K=8K=8 and 0.057 at K=20K=20; the interval excludes zero at both, also when images are resampled. Adding a personal strictness to the oracle, which then uses both the person’s answers and everyone’s votes on the test scene, gives the reference row of Table 1. § 5.7 shows who gains most.

Predicted properties help. PassCAL is better than the embedding alone by 0.004 at K=8K=8 (better for 73% of people) and 0.006 at K=20K=20 (74%), in all 20 partitions. With images resampled the difference stays clear of zero at K=8K=8 ([0.0003, 0.008]) and just reaches it at $K=20. Attributes parsed from WalkGPT [43], a VLM for pedestrian navigation, make the embedding worse instead (0.490 at K=8K=8; PassCAL better by 0.011, two-way CI [0.003, 0.018]): it names curbs, vehicles and vegetation in all fifty scenes and never a curb ramp or surface problem. Audited properties are ahead by 0.016 and 0.019; Sec. 5.6 locates this gap.

A per-person model does not suffice. The Linear-Hidden model of Ryu and Yanaka [38], fit on the same eight answers, reaches a Brier score of 0.550 on the 4096-dimensional VLM state and 0.539 on our 26-dimensional representation, against 0.480 PassCAL on the same items (better for 93% of people), and its per-person rank correlation with the answers is 0.10 against 0.31. With eight answers a person cannot be learned from scratch; what can be learned is how the person departs from a model of everyone else. The personal adapters of Sec. 5.4 are such shrinkage regressions, a hierarchical model with the group model as prior.

Prompting the same VLM does not personalise. On five partitions with identical people, calibration answers and test images, the VLM with eight in-context answers has a raw Brier score of 0.766: its answer probabilities are poorly calibrated. We therefore recalibrate it with a per-aid multinomial logistic map fit on eight-answer prompts built from training people and training images of the same fold, so that the map matches the test condition and uses no test person or test image. This gives 0.562, against 0.472 for PassCal on the same partitions (−0.090, CI [−0.105, −0.076], p=3×10−19p=3\times10^{-19}, better for 84% of people). The recalibrated VLM with eight answers is worse than our model with none (0.548, p=0.02p=0.02); the 32B model gives 0.563. Neither more examples nor a summary of the person’s answers changes this: with twenty answers in the prompt the recalibrated VLM scores 0.546 against 0.454 for PassCal (p=5×10−21p=5\times10^{-21}, better for 89% of people), and stating how often the person answered yes, unsure and no gives 0.556 against 0.472 (supplement). On these judgments the VLM does not extract the personal signal from the examples, although its hidden state describes the scene well enough for PassCal to do so; Ryu and Yanaka [38] likewise find a ten-shot prompt far below per-user heads for personal aesthetics.

Statements with controlled risk

Table 2 and Figure 4 report the decisive share. The realised frequency of a wrong PASSABLE per scene is 0.035 at α=0.05\alpha=0.05 and 0.085 at α=0.10\alpha=0.10, within the bound; among the scenes that receive PASSABLE at α=0.05\alpha=0.05, 13.6% are answered no, a ratio the bound does not control (§ 6). Without calibration answers the system commits on 27% of scenes, and with eight on 42% (+15.3 points paired, p=6×10−13p=6\times10^{-13}, more decisive for 74% of people). The comparison with the crowd oracle is clear here: +10.4 points at K=8K=8 (p=2×10−6p=2\times10^{-6}) and +14.7 at K=20K=20 (p=10−8p=10^{-8}). PassCal exceeds the embedding alone by 0.9 points at K=8K=8 (p=3×10−4p=3\times10^{-4}), and audited properties would add a further 2.2.

Modelα=0.05\alpha=0.05 K=0K=0820α=0.10\alpha=0.10 K=0K=0820
Embedding only27.241.045.857.075.378.5
PassCal26.641.946.257.676.079.4
Audited properties29.744.149.061.079.884.1
Crowd oracle31.566.2

Table 2. Decisive share (%): scenes that receive PASSABLE or BLOCKED when the expected frequency of each wrong statement is bounded at α\alpha (20 repetitions).

Figure 4. Decisive share at a 5% bound per statement type, against the number of calibration answers.

Which personal parameters matter?

Table 3 compares ways of spending the KK answers. Strictness adds one personal offset bpb_p to Eq. (1). Property weights fits bpb_p and δp\delta_p. Response calibration applies Eq. (2) to the group prediction and has no image-dependent personal parameter. PassCal combines property weights with response calibration. Each has its priors tuned on training people.

Personal adapterBrier ↓\downarrow K=3K=3820
Response calibration, no image0.5480.5210.503
Strictness0.5170.4930.480
Property weights0.5140.4870.469
Response calibration0.5130.4830.464
PassCal (weights + response)0.5120.4790.457

Table 3. Personal parameterisations on the same people, answers and test images. The first row fits the response calibration without the scene; the others share the same group model of the scene.

Personal calibration needs the scene. Fit to the aid’s answer distribution without the image, the response calibration scores 0.521 at K=8K=8, worse than PassCal by 0.044 (two-way CI [0.025, 0.063], 85% of people). On top of the scene model, a person’s response style is the first thing the answers reveal: strictness alone recovers 0.057 of the 0.070 reduction at K=8K=8, and response calibration, which also adapts the use of “unsure” and the confidence, adds 0.010 at K=8K=8 (p=2×10−4p=2\times10^{-4}) and 0.016 at K=20K=20 (p=7×10−10p=7\times10^{-10}).

Property weights add to it. Combining the two is better than response calibration alone by 0.004 at K=8K=8 (p=0.001p=0.001) and 0.007 at K=20K=20 (p=10−7p=10^{-7}, better for 70% of people, in all 20 partitions), better than property weights alone by 0.008 (p=0.03p=0.03) and 0.012 (p=6×10−4p=6\times10^{-4}), and makes more decisive statements than response calibration (+1.1 and +1.6 points, p≤10−5p\leq10^{-5}). The weights need more than three answers to show an effect. The decisive shares of all adapters are in the supplement.

Are the properties the right personal axes? Moving the personal weights to the first ten principal axes of the embedding, with the group model unchanged, is worse by 0.003 at K=8K=8 (p=0.002p=0.002) and 0.005 at K=20K=20 (p=10−5p=10^{-5}, 68% of people). Moving them to ten random directions of the same 26-dimensional space, which mix properties and embedding, matches PassCAL (difference at most 0.003, p≥0.1p \ge0.1). The personal signal therefore lies in the joint span of the representation, which the leading embedding directions alone do not cover; the properties are one sufficient choice of axes, and the one whose parameters can be read, not the only one.

Does it depend on the encoder?

We replace the scene embedding with each of sixteen frozen encoders, keeping the detector and the personal model fixed (Table 4): eight vision encoders [12, 31, 36, 45] through their pooled output, and eight VLMs [2, 4, 5, 26] through the answer-position state of § 4.2 (the supplement lists the checkpoints).

EncoderBrier, K=20K=20 Emb.+Prop.Δ\DeltaDecisive, K=20K=20 Emb.+Prop.
Vision encoders (pooled output)
CLIP ViT-B/320.4790.470−0.00941.643.6
CLIP ViT-B/160.4930.478−0.01539.542.6
CLIP ViT-L/140.4930.477−0.01538.941.9
DINOv2-B0.4940.484−0.01039.741.6
DINOv2-L0.5050.494−0.01138.040.2
SigLIP2-B0.4690.462−0.00743.545.7
SigLIP2-So400m0.4990.489−0.01038.941.1
ViT-B/16 (sup.)0.5060.489−0.01837.240.6
Vision-language models (answer-position state)
LLaVA-1.5 7B0.4790.470−0.00941.943.9
Qwen2.5-VL 7B0.4800.471−0.00943.144.4
Cosmos-Reason1 7B0.4760.466−0.01042.444.7
Qwen3.5 9B0.4830.468−0.01542.144.6
Qwen3.6 27B0.4700.464−0.00743.945.1
Qwen3.8 27B0.4740.462−0.01242.244.7
Qwen3-VL 32B0.4630.457−0.00744.746.2
Qwen3-VL 8B (ours)0.4640.457−0.00645.846.2

Table 4. Sixteen frozen encoders as the scene embedding, without (Emb.) and with (+Prop.) the predicted properties, over 20 partitions. Every Δ\Delta is significant (p≤2×10−7p \le2 \times10^{-7}, people paired). Decisive share in % at α=0.05\alpha=0.05.

Properties help every encoder. Adding the predicted properties improves all sixteen, by 0.006 to 0.018 Brier at K=20K=20 and by 0.5 to 3.4 points of decisive share. The audit-supervised properties consistently add information that neither general-purpose vision embeddings nor the states of large VLMs expose to a model fit on fifty scenes.

Which encoder. Qwen3-VL gives the best scores, and size is not what matters: the 8B model equals the 32B one. The other recent VLMs and the best vision encoder (SigLIP2-B) follow within 0.007, and older VLMs are at the level of the vision encoders. We adopted Qwen3-VL-8B from this comparison. Choosing the encoder within each fold from the training-only tuning loss gives 0.486 at K=8K=8, with a Qwen3-VL model chosen in 68 of 100 folds; the supplement also repeats the analyses with a DINOv2 encoder fixed beforehand. Controls in the supplement show that the useful information is in the state, not in the VLM’s answer: adding its answer probability to a DINOv2 embedding changes nothing, and the mean of the image tokens or a neutral prompt is worse.

What must perception get right?

The gap is type and severity. Table 5 replaces one block of predicted properties at a time with the audited values. Audited severity gains 0.012 at K=20K=20 and audited type 0.009, the two together account for the whole gap, and audited tags add nothing. On the survey images the detector identifies the type in 72% of cases, with recall of 84% for surface problems and 75% for obstacles but 60% for curb ramps (3 of 5) and 33% for missing curb ramps (2 of 6). The missing curb ramp is also the property on which people differ most (Sec. 5.4), so the class the detector recognises worst is the one that matters most for personalisation. Tag AUCs range from 0.76 to 0.96. Predicted severity correlates with the audited value at r=0.32r=0.32; two Project Sidewalk auditors who label the same object agree on its severity within-type rr between 0.11 and 0.44 (25,157 pairs), so severity labels are themselves noisy. Linear probes on the VLM state recover these properties less well than the detector (type accuracy 56%), which is why the two sources are complementary.

Properties in the modelK=8K=8K=20K=20pp
All predicted (PassCAL)0.4790.457
Audited tags0.4800.4570.14
Audited severity0.4670.4456×10−66 \times10^{-6}
Audited type0.4740.4494×10−54 \times10^{-5}
Audited type and severity0.4630.4381×10−101 \times10^{-10}
All audited0.4630.4382×10−82 \times10^{-8}

Table 5. Oracle study (Brier): one block of predicted properties is replaced by its audited value. pp: paired test against PassCAL at K=20K=20.

Training data and geography. Growing the detector’s pool from 8.4k to 14.9k scenes raises type accuracy on held-out crops from 0.73–0.75 to 0.79–0.80 and lowers the personal Brier at K=20K=20 from 0.460 to 0.457 (p=7×10−9p=7 \times10^{-9}); 18.1k scenes add nothing.

An unseen barrier type. To test transfer beyond the scene distribution, we withhold one Project Sidewalk label type at a time: the group model is fit on the other three types and evaluated on the withheld one, with people split as before. Eight answers still lower the Brier score from 0.547 to 0.504 and twenty to 0.488 (p≤10−3p \le10^{-3}), and the properties improve on the embedding alone by 0.006 (p=2×10−10p=2 \times10^{-10}, 79% of people). The supplement compares the adapters and the crowd oracle on this split.

Who benefits

Gains by aid (supplement). With eight answers the gain is largest for walker and cane users (0.115 and 0.096, p≤2×10−5p \le2 \times10^{-5}) and significant for manual wheelchair users (p=0.002p=0.002), who are already below the crowd oracle. Motorized wheelchair and scooter users need twenty answers to fall below it; the gain is then significant for the former (p=2×10−4p=2 \times10^{-4}) but not for the 22 scooter users (p=0.07p=0.07).

People who disagree with the crowd gain more. By distance of the eight calibration answers from others’ average answer, the Brier gain is 0.039 in the two quartiles closest to the crowd and 0.070 and 0.157 in the two farthest (supplement): those a population model serves worst gain most.

A second population: rater strictness

The survey has 135 people. To test whether few-shot personal calibration is specific to it, we use a second population in the same visual domain. Every Project Sidewalk label carries the identifier of the volunteer auditor who placed it and the severity that auditor gave, an individual and subjective judgment of a labelled object. The task mirrors § 5.1: given KK ratings from an auditor the model has not seen, predict the severity they give to other scenes. We use the same ordinal model with a personal strictness and five folds over auditors.

With images. For the 95 auditors with at least 30 ratings in our pool, the input is the embedding of the label crop and the label type. No panorama of a test auditor is used to fit the group model, and an auditor’s calibration and test ratings come from different panoramas. Eight ratings improve the mean Brier by 0.034 (CI [0.018, 0.052], p=0.026p=0.026) and twenty by 0.043 (p=8×10−4p=8 \times10^{-4}; supplement). Here the personal signal is mostly one number, the auditor’s strictness, and a prompt can carry it: Qwen3-VL-8B with the same eight ratings in context reaches 0.615 after recalibration, level with the fitted model on the same items (0.613, p=0.6p=0.6). On the mobility judgments, where the answer depends on the scene and the person together, it does not (§ 5).

Without images, at scale. Personal strictness can be measured on far more auditors if the image is dropped. With label type and city as the only inputs, on 996 auditors and 438k ratings, one rating already helps (−0.009, p=10−23p=10^{-23}), eight give −0.040 (better for 74% of auditors) and twenty −0.054. Auditors are volunteers and not mobility-aid users, so both experiments support the stability of personal strictness, not claims about accessibility needs.

Limitations

Scale and target. The binding constraint is the 50 scenes, the only public dataset of individual passability judgments by mobility-aid users that we know of. Repeated partitions cannot add visual diversity the pool does not have; the two-way bootstrap and the held-out barrier type measure this limit. Some properties, missing curb ramps among them, appear in as few as 6 of the 50 images. The 135 people are adequate for the pooled comparisons; the smallest aid group (22 scooter users) is not, and its weaker significance reflects that. The encoder was adopted after comparing sixteen on this survey; nested selection costs 0.007 (Sec. 5.5). Judgments concern photographs, not traversals, and the detector is trained on audits from the survey cities.

The guarantee is marginal. The bound of § 4.4 holds on average over people and counts errors over all of a person’s scenes. It does not control the error rate among the scenes that receive PASSABLE, which ranges from 11% for cane and walker users to 29% for scooter users at α=0.05\alpha=0.05 (supplement). Thresholds calibrated per aid restore the bound within each aid, but need at least 1/α−11/\alpha-1 calibration people per group, more than the smaller groups have at α=0.05\alpha=0.05.

Use. PassCal is not a navigation safety system: the error among PASSABLE statements is not bounded, and a deployment would need conservative, group-specific thresholds calibrated on more people than any public dataset has. Personal parameters reveal a person’s aid and tolerance and should stay on the person’s device. We use only the public release of Li et al., whose survey collected these judgments for research on personalised maps, and public Project Sidewalk labels, with auditor identifiers anonymised.

Conclusion

A sidewalk predictor should specify whose judgment it predicts. PassCal adapts audit-supervised scene properties to a person from a few answers. On new people and scenes, eight answers cut the Brier score by 0.070 and raise the decisive share at a 5% risk bound from 27% to 42%.

References

  1. [1]Anastasi­os N. Angelopoulos, Stephen Bates, Adam Fisch, Lihua Lei, and Tal Schuster. Conformal risk control. In International Conference on Learning Representations (ICLR), 2024. 2, 3, 4
  2. [2]Alisson Azzolini, Hannah Brandon, Prithvijit Chattopadhyay, Huayu Chen, Jinju Chu, Yin Cui, Jenna Diamond, Yifan Ding, Francesco Ferroni, Rama Govindaraju, et al. Cosmos-Reason1: From physical common sense to embodied reasoning. arXiv preprint arXiv:2503.15558, 2025. 7arxiv.org/abs/2503.15558
  3. [3]Jiamu Bai, Jiaming Hu, Yanhong Wu, and Zellux Wang. Personalize at test time: Learning user preferences for image generation. arXiv preprint arXiv:2610.09015, 2026. 3arxiv.org/abs/2610.09015
  4. [4]Shuai Bai, Yuxuan Cai, Ruize Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-VL technical report. arXiv preprint arXiv:2511.21631, 2025. 3, 7arxiv.org/abs/2511.21631
  5. [5]Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2.5-VL technical report. arXiv preprint arXiv:2502.13923, 2025. 7
  6. [6]Nicholas Bolten, Sumit Mukherjee, Veronika Sipeeva, Anissa Tanweer, and Anat Caspi. A pedestrian-centered data approach for equitable access to urban infrastructure environments. IBM Journal of Research and Development, 61 (6):10:1–10:12, 2017. 1, 2DOI
  7. [7]Kushal Chauhan, Rishabh Tiwari, Jan Freyberg, Pradeep Shenoy, and Krishnamurthy Dvijotham. Interactive concept bottleneck models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5948–5955, 2023. 3
  8. [8]Katherine M. Collins, Umang Bhatt, and Adrian Weller. Eliciting and learning with soft labels from every annotator. In Proceedings of the AAAI Conference on Human Computation and Crowdsourcing (HCOMP), pages 40–52, 2022. 3
  9. [9]R. J. P. Damaceno, L. Ferreira, F. Miranda, M. Hosseini, and R. M. Cesar Jr. SideSeeing: A multimodal dataset and collection of tools for sidewalk assessment. arXiv preprint arXiv:2407.06464, 2024. 2arxiv.org/abs/2407.06464
  10. [10]Molloud Damandeh and Me ad Saberi. Walkable to whom? Capturing subjective variability in walkability perception using multimodal deep learning. arXiv preprint arXiv:2608.06934, 2026. 3arxiv.org/abs/2608.06934
  11. [11]Aida Mostafazadeh Davani, Mark Díaz, and Vinodkumar Prabhakaran. Dealing with disagreements: Looking beyond the majority vote in subjective annotations. Transactions of the Association for Computational Linguistics, 10:92–110, 2022. 3arxiv.org/abs/2110.05719
  12. [12]Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 7
  13. [13]Abhimanyu Dubey, Nikhil Naik, Devi Parikh, Ramesh Raskar, and César A. Hidalgo. Deep learning the city: Quantifying urban perception at a global scale. In European Conference on Computer Vision (ECCV), 2016. 3arxiv.org/abs/1608.01769
  14. [14]Lohic Fotio Tiotsop, Antonio Servetti, Peter Pocta, Glenn Van Wallendael, Marcus Barkowsky, and Enrico Masala. Multiple image distortion DNN modeling individual subject quality assessment. ACM Transactions on Multimedia Computing, Communications, and Applications, 20(8):1–27, 2024. 3
  15. [15]Samuel Goree, Weslie Khoo, and David J. Crandall. Correct for whom? subjectivity and the evaluation of personalized image aesthetics assessment models. In Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 3
  16. [16]Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. In Proceedings of the 34th International Conference on Machine Learning (ICML), pages 1321–1330. PMLR, 2017. 3, 4arxiv.org/abs/1706.04599
  17. [17]Kotaro Hara, Vicki Lee, and Jon Froehlich. Combining crowdsourcing and Google Street View to identify street-level accessibility problems. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, page 631–640, New York, NY, USA, 2013. Association for Computing Machinery. 2DOI
  18. [18]Kotaro Hara, Jin Sun, Robert Moore, David Jacobs, and Jon Froehlich. To me: Detecting curb ramps in Google Street View using crowdsourcing, computer vision, and machine learning. In Proceedings of the 27th Annual ACM Symposium on User Interface Software and Technology, page 189–204, New York, NY, USA, 2014. Association for Computing Machinery. 2DOI
  19. [19]Heidelberg Institute for Geoinformation Technology. openrouteservice: Wheelchair profile and tag filtering documentation, 2026. Accessed: 2026-05-20. 1
  20. [20]Maryam Hosseini, Fabio Miranda, Jian zhe Lin, and Claudio T. Silva. CitySurfaces: City-scale semantic segmentation of sidewalk materials. Sustainable Cities and Society, 79:103630, 2022. 2DOI
  21. [21]Maryam Hosseini, Andres Sevtsuk, Fabio Miranda, Roberto M. Cesar Jr, and Claudio T. Silva. Mapping the walk: A scalable computer vision approach for generating sidewalk network datasets from aerial imagery. Computers, Environment and Urban Systems, 101:101950, 2023. 2DOI
  22. [22]William Huang, Xia Su, Jon E. Froehlich, and Yang Zhang. Accessibility scout: Personalized accessibility scans of built environments. In Proceedings of the 38th Annual ACM Symposium on User Interface Software and Technology, pages 1–18, 2025. 2DOI
  23. [23]Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International Conference on Machine Learning (ICML), 2020. 3arxiv.org/abs/2007.04612
  24. [24]Chu Li, Rock Yuren Pang, Delphine Labbé, Yochai Eisenberg, Maryam Hosseini, and Jon E. Froehlich. Accessibility for whom? Perceptions of mobility barriers across disability groups and implications for designing personalized maps. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, pages 1–19, 2025. 2, 3DOI
  25. [25]Seung Jae Lieu, Diego Morra, Chiara Cadoni, Wonseop Song, Martina Mazzarello, and Carlo Ratti. Can VLMs reliably assess sidewalk accessibility attributes from pedestrian-level imagery? arXiv preprint arXiv:2609.17882, 2026. 2, 3arxiv.org/abs/2609.17882
  26. [26]Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 26296–26306, 2024. 7arxiv.org/abs/2310.03744
  27. [27]Xinlei Liu, Kevin Wu, Minchu Kulkarni, Michael Saugstad, Peyton Anton Rapo, Jeremy Freiburger, Maryam Hosseini, Chu Li, and Jon E. Froehlich. Towards fine-grained sidewalk accessibility assessment with deep learning: Initial benchmarks and an open dataset. In Proceedings of the 26th International ACM SIGACCESS Conference on Computers and Accessibility, New York, NY, USA, 2024. Association for Computing Machinery. 2
  28. [28]Kelly Mack, Emma McDonnell, Dhruv Jain, Lucy Lu Wang, Jon E. Froehlich, and Leah Findlater. What do we mean by “accessibility research”? A literature survey of accessibility papers in CHI and ASSETS from 1994 to 2019. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems, pages 1–18, 2021. 2
  29. [29]Peter McCullagh. Regression models for ordinal data. Journal of the Royal Statistical Society: Series B, 42(2):109–142, 1980. 4DOI
  30. [30]John S. O’Meara, Jared Hwang, Zeyu Wang, Michael Saugstad, and Jon E. Froehlich. RampNet: A two-stage pipeline for bootstrapping curb ramp detection in streetscape images from open government metadata. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 6715–6724, 2025. 2
  31. [31]Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jégou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3, 7arxiv.org/abs/2304.07193
  32. [32]Sangwoo Park, Kfir M. Cohen, and Osvaldo Simeone. Few-shot calibration of set predictors via meta-learned cross-validation-based conformal prediction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(1):280–291, 2024. 3DOI
  33. [33]Yi-Hao Peng, Jeffrey P. Bigham, and Jason Wu. Efficient personalization of generative user interfaces. arXiv preprint arXiv:2604.09876, 2026. 3arxiv.org/abs/2604.09876
  34. [34]Joshua C. Peterson, Ruairidh M. Battleday, Thomas L. Griffiths, and Olga Russakovsky. Human uncertainty makes classification more robust. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9616–9625, 2019. 3arxiv.org/abs/1908.07086
  35. [35]Matias Quintana, Youlong Gu, Xiucheng Liang, Yujun Hou, Koichi Ito, Yihan Zhu, Mahmoud Abdelrahman, and Filip Biljecki. Global urban visual perception varies across demographics and personalities. Nature Cities, 2(11):1092–1106, 2025. 3DOI
  36. [36]Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. 7arxiv.org/abs/2103.00020
  37. [37]Jian Ren, Xiaohui Shen, Zhe Lin, Radomir Mech, and David J. Foran. Personalized image aesthetics. In IEEE International Conference on Computer Vision (ICCV), 2017. 3DOI
  38. [38]Koki Ryu and Hitomi Yanaka. What do vision-language models encode for personalized image aesthetics assessment? In Findings of the Association for Computational Linguistics: ACL 2026, 2026. arXiv:2604.11374. 2, 3, 5, 6arxiv.org/abs/2604.11374
  39. [39]Manaswi Saha, Michael Saugstad, Hanuma Teja Maddali, Aileen Zeng, Ryan Holland, Steven Bower, Aditya Dash, Sage Chen, Anthony Li, Kotaro Hara, and Jon Froehlich. Project Sidewalk: A web-based crowdsourcing tool for collecting sidewalk accessibility data at scale. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems, page 1–14, New York, NY, USA, 2019. Association for Computing Machinery. 2
  40. [40]Shaswati Saha, Lauren Selingo, Emily Olejniczak, Hanna Noyce, Vaskar Raychaudhury, Roger O. Smith, and Md Osman Gani. MyPath: Accessible routing for wheelchair users. In RESNA Annual Conference, 2022. 1, 2DOI
  41. [41]Murilo Salem, Luísa Böhm, Daniel Pontes, and Anderson Ferrugem. Hierarchical group-conditional conformal risk control for selective prediction in language models. arXiv preprint arXiv:2607.24562, 2026. 3arxiv.org/abs/2607.24562
  42. [42]David Steinmann, Wolfgang Stammer, Felix Friedrich, and Kristian Kersting. Learning to intervene on concept bottlenecks. In Proceedings of the 41st International Conference on Machine Learning, pages 46556–46571, 2024. 3arxiv.org/abs/2308.13453
  43. [43]Rafi Ibn Sultan, Hui Zhu, Xiangyu Zhou, Chengyin Li, Prashant Khanduri, Marco Brocanelli, and Dongxiao Zhu. WalkGPT: Grounded vision-language conversation with depth-aware segmentation for pedestrian navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026. 2, 5
  44. [44]Benjamin Tannert and Johannes Schöning. Disabled, but at what cost? An examination of wheelchair routing algorithms. In Proceedings of the 20th International Conference on Human-Computer Interaction with Mobile Devices and Services, pages 1–7, 2018. 1, 2DOI
  45. [45]Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Alabdulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, Olivier Hénaff, Jeremiah Harmsen, Andreas Steiner, and Xiaohua Zhai. SigLIP 2: Multilingual vision-language encoders with improved semantic understanding, localization, and dense features. arXiv preprint arXiv:2502.14786, 2025. 7
  46. [46]Dongdong Wang, Alina Hagen, Isabelle Gatmaltain, Hao Zhou, Yiwen Dong, Shabboo Valipoor, Vivian W. H. Wong, and Lingyao Li. Do VLMs see what sensors feel? A scalable expert-guided design for wheelchair accessibility assessment from street view. arXiv preprint arXiv:2606.07642, 2026. 2arxiv.org/abs/2606.07642
  47. [47]Galen Weld, Esther Jang, Anthony Li, Aileen Zeng, Kurtis Heimerl, and Jon E. Froehlich. Deep learning for automatically detecting sidewalk accessibility problems using streetscape imagery. In Proceedings of the 21st International ACM SIGACCESS Conference on Computers and Accessibility, page 196–209, New York, NY, USA, 2019. Association for Computing Machinery. 2DOI
  48. [48]Ruoyu Xue, Hieu Le, Jingyi Xu, Sounak Mondal, Abe Leite, Gregory Zelinsky, Minh Hoai, and Dimitris Samaras. Personalized image descriptions from attention sequences. arXiv preprint arXiv:2512.06662, 2025. 3arxiv.org/abs/2512.06662
  49. [49]Ruoyu Xue, Jingyi Xu, Sounak Mondal, Hieu Le, Gregory Zelinsky, Minh Hoai, and Dimitris Samaras. Few-shot personalized scanpath prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 3arxiv.org/abs/2504.05499
  50. [50]Yuzhe Yang, Liwu Xu, Leida Li, Nan Qie, Yaqian Li, Peng Zhang, and Yandong Guo. Personalized image aesthetics assessment with rich attributes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3arxiv.org/abs/2203.16754
  51. [51]Zhichao Yang, Tianjiao Gu, Zhixianhe Zhang, Xiangfei Sheng, Pengfei Chen, and Leida Li. Personalized image aesthetic assessment via preference-rich sample mining and cohort merging. arXiv preprint arXiv:2607.15752, 2026. Authors report acceptance at ACM MM 2026. 3arxiv.org/abs/2607.15752
  52. [52]Mert Yuksekgonul, Maggie Wang, and James Zou. Post-hoc concept bottleneck models. In International Conference on Learning Representations (ICLR), 2023. 3arxiv.org/abs/2205.15480
  53. [53]Jooyeol Yun and Jaegul Choo. Scaling up personalized image aesthetic assessment via task vector customization. In European Conference on Computer Vision (ECCV), 2024. 3arxiv.org/abs/2407.07176

Paper details

vlnc:W_cyxaymvk
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