Pith. sign in

REVIEW 3 major objections 5 minor 52 references

How Well Do Vision--Language Models Understand Cities? A Comparative Study on Spatial Reasoning from Street-View Images

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Fine-tuning vision-language models on a synthetic street-view question dataset substantially improves their spatial reasoning, especially on negation, counterfactuals, and depth judgments.

desk verdict Useful synthetic-data study with an honest but real circularity problem in the evaluation; the gains are real relative to the authors' rules, not yet proven against human spatial reasoning. read the letter →

arxiv 2508.21565 v1 pith:KEOFAUGJ submitted 2025-08-29 cs.CV

Juneyoung Ro , Namwoo Kim , Yoonjin Yoon This is my paper · ORCID
classification cs.CV
keywords vision-languagemodelsspatialreasoningurbansceneunderstandingsyntheticvisualquestionansweringchain-of-thoughtsupervisionstreet-viewimagerynegationcounterfactual
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper asks how well general-purpose vision-language models can reason about street-level urban scenes and whether targeted fine-tuning can close the gap. The authors build a synthetic dataset of about 280,000 question-answer pairs from 50,000 street-view images, using pretrained segmentation, object detection, and depth models to produce structured scene metadata. Questions range from simple perceptual judgments to negation, counterfactuals, and multi-hop comparisons, and each answer is expanded into a chain-of-thought rationale. They find that fine-tuning on this synthetic supervision substantially boosts spatial reasoning, with the largest gains on exactly the question types that trip up zero-shot models: negation, counterfactuals, and depth-based judgments. If correct, the result suggests that a modest amount of in-domain synthetic supervision can give general-purpose vision-language models a meaningful portion of the fine-grained spatial understanding that urban scene analysis requires.

What carries the argument

The machinery is a metadata-driven QA pipeline. Semantic segmentation yields the pixel proportions of greenery, sky, and buildings; object detection yields counts and locations; monocular depth estimation yields depth range, variance, and closest-object information. These are fused into a structured metadata record for each image, and deterministic, question-type-specific rules turn that record into base question-answer pairs. A text-generating language model then rewrites each short answer into a step-by-step chain-of-thought rationale that stays grounded in the same metadata, providing the reasoning supervision used for fine-tuning.

What would settle it

Hand-label the ground truth for a random sample of the 280,000 question-answer pairs, including the object-count cases the paper's own human check flagged, then re-score the zero-shot and fine-tuned models against those human labels; if the model ranking or the size of the fine-tuning gains changes materially, the reported improvements are an artifact of the synthetic labels.

Watch

Extended reading notes

Core claim

The central claim, on the paper's own terms, is that a synthetic, chain-of-thought-supervised question-answer dataset assembled from pretrained scene metadata is a practical path to adapt general-purpose vision-language models to fine-grained urban spatial reasoning. Fine-tuning BLIP-2 on this dataset raises its negation F1 from 0.52 to 0.91, a 75 percent gain, and its counterfactual F1 from 0.55 to 0.90, a 64 percent gain; depth-categorical F1 rises from 0.11 to 0.76, a 591 percent gain, and depth-closest-object F1 from 0.11 to 0.67, a 509 percent gain. The gains are not uniform: object presence improves only slightly, and some simple perceptual tasks, notably proportion binary, object co-occurrence, and layout binary, fall after fine-tuning, which the paper attributes to catastrophic forgetting or distribution mismatch. The paper also documents distinct model profiles: LLaVA-1.5 is the strongest zero-shot reasoner, BLIP-2 benefits the most from fine-tuning, and InstructBLIP shows strength on counterfactuals but wider perceptual degradation.

Load-bearing premise

The answers used for both training and evaluation are generated from automated estimates of depths, object counts, and scene proportions rather than from direct human verification, so systematic errors in those estimates would systematically distort both what the models learn and what the tests measure.

Editorial extensions

If this is right

  • Fine-tuning on a few thousand in-domain synthetic examples can take a lightweight model like BLIP-2 from weak on negation and depth to strong, suggesting model size is not the main barrier to urban spatial reasoning.
  • The largest gains occur exactly on negation, counterfactuals, and depth, so these capabilities appear to be more data-limited than architecturally limited in current vision-language models.
  • Simple perceptual tasks can regress after fine-tuning, so domain adaptation should be paired with rehearsal data, multi-task balancing, or selective freezing to preserve perceptual competence.
  • Synthetic metadata-grounded question-answer generation is a scalable and reproducible route to evaluation and supervision for specialized visual domains, avoiding large-scale human annotation.
  • Zero-shot strengths differ by model, so deployment choices should weigh a model's baseline robustness against how much it gains from targeted fine-tuning.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension of this result is to vary the amount of chain-of-thought supervision versus plain answer supervision during fine-tuning, which would isolate how much of the gain comes from the reasoning traces rather than from the answer labels alone.
  • The same metadata-driven supervision approach could be transplanted to other structured scene domains where pretrained extractors exist, such as indoor navigation, rural land-use assessment, or medical imagery, without requiring new human-annotated benchmarks.
  • If the observed negation and counterfactual gains replicate in other settings, it would suggest that these reasoning gaps are distributional rather than fundamental, and that larger or more domain-specific pretraining corpora could recover part of the fine-tuning effect.
  • Because perceptual tasks degraded in some models, a practical deployment implication is that a fine-tuned model should not be treated as a drop-in replacement for the zero-shot model across all question types.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. This paper presents a synthetic VQA dataset for fine-grained spatial reasoning in urban street-view images. The dataset is generated from metadata extracted by pretrained perception models (SegFormer for segmentation, DETR for object detection, MiDaS for depth) applied to street-view images from five cities, with deterministic rules producing base QA pairs and Gemini 1.5-Flash expanding them into chain-of-thought rationales. The authors evaluate BLIP-2, InstructBLIP, and LLaVA-1.5 in zero-shot and fine-tuned settings, reporting that CoT fine-tuning on the synthetic dataset substantially improves accuracy, especially on negation, counterfactual, and depth-related questions. A human validation of 500 QA pairs is used to assess metadata and CoT quality.

Significance. If the central claim survives an independent evaluation, the paper would be a useful contribution: it introduces a large-scale, reproducible synthetic urban spatial reasoning dataset, an open pipeline, and a comparative study of three open-source VLMs. The 280K QA pairs and public code are valuable resources, and the idea of using structured metadata to supervise chain-of-thought reasoning is timely. However, the current evidence is insufficient because the evaluation is internal: test labels are generated by the same deterministic rules over the same perception-pipeline outputs that produce the training targets. The reported human validation covers only 0.18% of the dataset and checks plausibility rather than exact QA correctness. These issues bear directly on the paper's headline claim that fine-tuning improves spatial reasoning, so the contribution cannot yet be taken as established.

major comments (3)
  1. [§3.2–§3.3 and §4.2 (Tables 2–3)] The central conclusion rests on a circular evaluation. The test-set answers are produced by the same deterministic rules over the same SegFormer/DETR/MiDaS metadata that generates the training targets; a model can therefore 'improve' by learning to emulate the perception pipeline's systematic errors rather than by reasoning about the image. The human validation in §3.3.3 is not sufficient to break this circularity: it covers only 500 of roughly 280,000 QA pairs (0.18%), uses binary plausibility judgments rather than exact QA correctness, and reports object-detection accuracy of 88%, leaving about one in eight object-related labels questionable. To support RQ2, the authors should add an independent test set with human-annotated answers, ideally stratified by question type, and report per-type human agreement.
  2. [§4.2.1, Tables 2 and 3] The headline gains are relative changes from very low zero-shot baselines and are not consistent across models. For example, BLIP-2's depth-categorical F1 rises from 0.11 to 0.76 (+591%), while LLaVA-1.5's same metric drops from 0.62 to 0.50 and InstructBLIP's depth-closest drops from 0.22 to 0.10. The paper should report absolute F1/MAE values with confidence intervals and explain why BLIP-2's zero-shot depth performance is substantially worse than the other models; as written, the 'substantial boost' claim is not robust across the model set.
  3. [§3.3.1 and Fig. 6] The ground-truth labels for compositional questions depend on ad hoc thresholds, for example the crowdedness threshold of five people used in the counterfactual question and the binary proportion cutoffs; no sensitivity analysis or human-consensus validation of these thresholds is provided. Figure 6 also shows a fine-tuned BLIP-2 counterfactual answer ('no') that contradicts the ground truth ('Yes'), suggesting that the rule or label, rather than the model, may be wrong in some cases. The authors should provide a sensitivity analysis over thresholds and validate the actual QA labels, not just the underlying metadata, on a human-annotated sample.
minor comments (5)
  1. [§4.1] The fine-tuning protocol uses 40 epochs but does not mention early stopping or validation-based checkpoint selection; the paper should report whether the results are from the final epoch or a selected checkpoint.
  2. [Table 3] Relative percentage changes such as +1350% for the top-entity task are misleading for metrics starting near zero; absolute differences or log-odds would be more informative.
  3. [§4.2.1] The phrase 'a few thousand in-domain examples' appears inconsistent with the 280K QA pairs described in §1; the authors should specify the actual number of examples used in the fine-tuning set, ideally per task type.
  4. [Fig. 6 and Table 2] The fine-tuned BLIP-2 counterfactual example in Figure 6 contradicts the aggregate F1 of 0.90 reported in Table 2; the authors should explain this discrepancy or add a discussion of failure cases and whether such examples were human-verified.
  5. [Table 2 caption and §3.2, Table 1] The caption 'Bold indicates the better performance' is ambiguous for MAE, where lower is better, and the depth range reported as 41.5 lacks units; MiDaS predictions are scale-ambiguous, so the normalization should be specified.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the synthetic dataset is explicitly pseudo-grounded, and the evaluation is a standard holdout test of learning a well-defined image-to-answer mapping, with the labeler-bias limitation disclosed by the authors.

full rationale

The paper's pipeline generates QA labels from segmentation (SegFormer), detection (DETR), and depth (MiDaS) predictions and uses the same generation procedure for both the training and test splits. This makes the benchmark internal and limits the external, human-level interpretation of the absolute scores, but it does not make the derivation circular. The test-set answers are produced by fixed deterministic rules applied to metadata, not by the fine-tuned model, not by the training labels, and not by any parameter fitted to the test set. At test time the model receives only the raw street-view image and must recover the scene attributes itself; the failure cases in Figure 6 (e.g., fine-tuned BLIP-2 answering 'no' to a crowdedness counterfactual whose metadata-derived ground truth is 'yes') show that agreement with the labels is not automatic. The authors explicitly describe the metadata as 'pseudo-ground truth' in Section 3.2 and report a human validation with imperfect accuracy in Section 3.3.3 (e.g., 88% for object detection, with over-counting), so the shared-labeler-bias concern is stated as a limitation rather than concealed. There are no load-bearing self-citations, no imported uniqueness theorems, and no fitted parameter renamed as a prediction. The central claim—that fine-tuning on this synthetic CoT-supervised distribution improves holdout performance on questions from that same distribution—is self-contained and does not reduce to its inputs by construction.

Assumptions & free parameters 2 free parameters · 2 assumptions · 0 invented entities

The central claim rests on the accuracy of pretrained perception models that generate the pseudo-ground truth, and on the quality of LLM-generated CoT traces; both are validated only on a small sample. The hand-chosen thresholds in the QA templates are design parameters rather than fitted constants.

free parameters (2)
  • crowdedness threshold for counterfactual questions = 5 people
    In the counterfactual example, crowdedness is defined as 5 or more people; this threshold is chosen by hand and determines the correct answer for a class of generated questions.
  • perception thresholds for binary proportion questions = e.g., GVI > 30%
    Binary 'dominance' questions apply thresholds drawn from prior literature (e.g., GVI > 30%), which are design choices baked into the QA generation rules.
assumptions (2)
  • domain assumption Pretrained perception models (SegFormer, DETR, MiDaS) produce outputs that approximate ground truth sufficiently for QA generation.
    The entire dataset and its correctness depend on these model outputs; the paper validates only 500 samples and reports object detection accuracy of 88%.
  • domain assumption The CoT traces generated by Gemini 1.5-Flash from the metadata and reasoning rules are consistent and plausible supervision.
    CoT answers are used as training targets; the paper's human validation found 98% consistency but 90% plausibility, so some traces are imperfect.

how reviews work

0 comments
Cite this review

Pith. "Pith review of How Well Do Vision--Language Models Understand Cities? A Comparative Study on Spatial Reasoning from Street-View Images." pith.science (2026). https://pith.science/paper/KEOFAUGJ

@misc{pith2026250821565,
  author       = {Pith},
  title        = {Pith review of: How Well Do Vision--Language Models Understand Cities? A Comparative Study on Spatial Reasoning from Street-View Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KEOFAUGJ}},
  note         = {Machine review of arXiv:2508.21565}
}
read the original abstract

Effectively understanding urban scenes requires fine-grained spatial reasoning about objects, layouts, and depth cues. However, how well current vision-language models (VLMs), pretrained on general scenes, transfer these abilities to urban domain remains underexplored. To address this gap, we conduct a comparative study of three off-the-shelf VLMs-BLIP-2, InstructBLIP, and LLaVA-1.5-evaluating both zero-shot performance and the effects of fine-tuning with a synthetic VQA dataset specific to urban scenes. We construct such dataset from segmentation, depth, and object detection predictions of street-view images, pairing each question with LLM-generated Chain-of-Thought (CoT) answers for step-by-step reasoning supervision. Results show that while VLMs perform reasonably well in zero-shot settings, fine-tuning with our synthetic CoT-supervised dataset substantially boosts performance, especially for challenging question types such as negation and counterfactuals. This study introduces urban spatial reasoning as a new challenge for VLMs and demonstrates synthetic dataset construction as a practical path for adapting general-purpose models to specialized domains.

Figures

Figures reproduced from arXiv: 2508.21565 by the authors.

Figure 1
Figure 1. Example illustrating a zero-shot model’s failure in han [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of our pipeline. (a) We take raw street-view images as inputs and extract three key predictions: segmentation masks, object detection results, and monocular depth masks. Each output contributes to the assembled metadata describing the street scene, including view factor proportions (greenery, sky, building), object counts, spatial layout distribution, depth complexity, and closest object information. (b) Th… view at source ↗
Figure 3
Figure 3. Example of a synthesized QA pair. The CoT answer faith [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Example question types from the QA generation pipeline. Each example illustrates a representative question, the corresponding [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Parameter efficiency of vision–language models based on [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Qualitative answers from three VLMs—InstructBLIP, BLIP-2, and LLaVA-1.5—under zero-shot and fine-tuned settings for [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 43 canonical work pages

  1. [1]

    Google street view: Capturing the world at street level

    Dragomir Anguelov, Carole Dulong, Daniel Filip, Christian Frueh, St ´ephane Lafon, Richard Lyon, Abhijit Ogale, Luc Vincent, and Josh Weaver. Google street view: Capturing the world at street level. Computer, 43(6):32–38, 2010. 3

    work page 2010

  2. [2]

    Evaluation methods for landscapes with greenery

    Yoji Aoki. Evaluation methods for landscapes with greenery. Landscape Research, 16:3–6, 1991. 4

    work page 1991

  3. [3]

    Browning, Jiaying Dong, Kuiran Zhang, Shuai Yuan, H¨useyin Ertan ˙Inan, Olivia McAnirlin, Dani T

    Mondira Bardhan, Fu Li, Mathew H.E.M. Browning, Jiaying Dong, Kuiran Zhang, Shuai Yuan, H¨useyin Ertan ˙Inan, Olivia McAnirlin, Dani T. Dagan, Allison Maynard, Katie Thurson, Fan Zhang, Ruoyu Wang, and Marco Helbich. From space to street: A systematic review of the associations between visi- ble greenery and bluespace in street view imagery and mental hea...

    work page 2024

  4. [4]

    The green window view index: automated multi-source visi- bility analysis for a multi-scale assessment of green window views

    Anna-Maria Bolte, Benjamin Niedermann, Thomas Kiste- mann, Jan-Henrik Haunert, Youness Dehbi, and Theo K ¨otter. The green window view index: automated multi-source visi- bility analysis for a multi-scale assessment of green window views. Landscape Ecology, 39(3):71, 2024. 4

    work page 2024

  5. [5]

    Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Gi- ancarlo Baldan, and Oscar Beijbom

    Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Gi- ancarlo Baldan, and Oscar Beijbom. nuscenes: A multi- modal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2

    2020

  6. [6]

    End-to- end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to- end object detection with transformers. 2020. 3

    work page 2020

  7. [7]

    Spatialvlm: Endowing vision-language models with spatial reasoning capabilities

    Boyuan Chen, Zhuo Xu, Sean Kirmani, Danny Driess, Pete Florence, Brian Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In CVPR, 2024. 2, 4

    work page 2024

  8. [8]

    Evaluating implied urban nature vital- ity in san francisco: An interdisciplinary approach combining census data, street view images, and social media analysis

    Mingze Chen, Yuxuan Cai, Shuying Guo, Ruilin Sun, Yang Song, and Xiwei Shen. Evaluating implied urban nature vital- ity in san francisco: An interdisciplinary approach combining census data, street view images, and social media analysis. Urban Forestry & Urban Greening, 95:128289, 2024. 1

    work page 2024

Show all 52 references
  1. [9]

    Visual chain- of-thought prompting for knowledge-based visual reasoning

    Zhenfang Chen, Qinhong Zhou, Yikang Shen, Yining Hong, Zhiqing Sun, Dan Gutfreund, and Chuang Gan. Visual chain- of-thought prompting for knowledge-based visual reasoning. Proceedings of the AAAI Conference on Artificial Intelligence, 38(2):1254–1262, 2024. 4

    2024

  2. [10]

    The cityscapes dataset for semantic urban scene understanding

    Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Re- hfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 2, 3

    2016

  3. [11]

    Instructblip: Towards general-purpose vision-language models with instruction tuning

    Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructblip: Towards general-purpose vision-language models with instruction tuning. In NIPS, 2023. 2

    2023

  4. [12]

    Feeling Nature: Measuring perceptions of biophilia across global biomes using visual AI,

    Fabio Duarte, Deborah Lefosse, Rohit Sanatani, Yuhao Kang, Arjan Timmeren, and Carlo Ratti. Feeling Nature: Measuring perceptions of biophilia across global biomes using visual AI,

  5. [13]

    Abhimanyu Dubey, Nikhil Naik, Devi Parikh, Ramesh Raskar, and C´esar A. Hidalgo. Deep Learning the City: Quantifying Urban Perception at a Global Scale. In Computer Vision – ECCV 2016, pages 196–212, Cham, 2016. Springer Interna- tional Publishing. 1

    2016

  6. [14]

    The processing of negation and polarity: An overview

    Carolin Dudschig, Barbara Kaup, Mingya Liu, and Juliane Schwab. The processing of negation and polarity: An overview. Journal of Psycholinguistic Research, 50(6):1199– 1213, 2021. 4

    2021

  7. [15]

    Vqa-lol: Visual question answering under the lens of logic

    Tejas Gokhale, Pratyay Banerjee, Chitta Baral, and Yezhou Yang. Vqa-lol: Visual question answering under the lens of logic. In European conference on computer vision, pages 379–

  8. [16]

    Fang-Ying Gong, Zhao-Cheng Zeng, Fan Zhang, Xiaojiang Li, Edward Ng, and Leslie K. Norford. Mapping sky, tree, and building view factors of street canyons in a high-density urban environment. Building and Environment, 134(1):155– 167, 2018. 3

    2018

  9. [17]

    Xuan He and Sylvia Y . He. Using mobile phone big data and street view images to explore the mismatch between walka- bility and walking behavior. Transportation Research Part A: Policy and Practice, 180:103946, 2024. 1

    2024

  10. [18]

    Natural Adversarial Examples, 2021

    Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Stein- hardt, and Dawn Song. Natural Adversarial Examples, 2021. arXiv:1907.07174 [cs]. 1

    2021 arXiv

  11. [19]

    Which street is hotter? street morphol- ogy may hold clues -thermal environment mapping based on street view imagery

    Yanjun Hu, Fengtao Qian, Hai Yan, Ariane Middel, Renwu Wu, Minghui Zhu, Qian Han, Kechun Zhao, Han Wang, Feng Shao, and Zhiyi Bao. Which street is hotter? street morphol- ogy may hold clues -thermal environment mapping based on street view imagery. Building and Environment, 26...

  12. [20]

    Hudson and Christopher D

    Drew A. Hudson and Christopher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR),

  13. [21]

    Lawrence Zitnick, and Ross Girshick

    Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C. Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recogniti...

    2017

  14. [22]

    Negated and misprimed probes for pretrained language models: Birds can talk, but cannot fly

    Nora Kassner and Hinrich Sch ¨utze. Negated and misprimed probes for pretrained language models: Birds can talk, but cannot fly. arXiv preprint arXiv:1911.03343, 2019. 4

    1911 arXiv

  15. [23]

    Large language models are zero-shot reasoners

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. In NIPS, 2022. 4

    2022

  16. [24]

    Understanding counterfactuality: A review of experimental evidence for the dual meaning of counterfactuals

    Eugenia Kulakova and Mante S Nieuwland. Understanding counterfactuality: A review of experimental evidence for the dual meaning of counterfactuals. Language and Linguistics compass, 10(2):49–65, 2016. 4

    2016

  17. [25]

    Blip- 2: Bootstrapping language-image pre-training with frozen im- age encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip- 2: Bootstrapping language-image pre-training with frozen im- age encoders and large language models. In ICML, 2023. 2

    2023

  18. [26]

    Assessing street-level urban greenery using google street view and a modified green view index

    Xiaojiang Li, Chuanrong Zhang, Weidong Li, Robert Ricard, Qingyan Meng, and Weixing Zhang. Assessing street-level urban greenery using google street view and a modified green view index. Urban Forestry & Urban Greening, 14:675–685,

  19. [27]

    Generalizing Vision-Language Mod- els to Novel Domains: A Comprehensive Survey, 2025

    Xinyao Li, Jingjing Li, Fengling Li, Lei Zhu, Yang Yang, and Heng Tao Shen. Generalizing Vision-Language Mod- els to Novel Domains: A Comprehensive Survey, 2025. arXiv:2506.18504 [cs]. 1

    2025 arXiv

  20. [28]

    Eyes can deceive: Benchmarking counterfactual rea- soning abilities of multi-modal large language models

    Yian Li, Wentao Tian, Yang Jiao, Jingjing Chen, and Yu-Gang Jiang. Eyes can deceive: Benchmarking counterfactual rea- soning abilities of multi-modal large language models. arXiv e-prints, pages arXiv–2404, 2024. 4

    2024

  21. [29]

    Evaluating human perception of building exteriors using street view imagery

    Xiucheng Liang, Jiat Hwee Chang, Song Gao, Tianhong Zhao, and Filip Biljecki. Evaluating human perception of building exteriors using street view imagery. Building and Environ- ment, 263:111875, 2024. 1

    2024

  22. [30]

    Improved baselines with visual instruction tuning

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR,

  23. [31]

    Efficacy of synthetic data as a benchmark

    Gaurav Maheshwari, Dmitry Ivanov, and Kevin El Haddad. Efficacy of synthetic data as a benchmark. arXiv preprint arXiv:2409.11968, 2024. 5

    2024 arXiv

  24. [32]

    Review of methods used to es- timate the sky view factor in urban street canyons

    Chunping Miao, Shuai Yu, Yuanman Hu, Huiwen Zhang, Xingyuan He, and Wei Chen. Review of methods used to es- timate the sky view factor in urban street canyons. Building and Environment, 168:106497, 2020. 4

    2020

  25. [33]

    Objective scoring of streetscape walkability related to leisure walking: Statistical modeling approach with semantic segmentation of google street view images

    Shohei Nagata, Tomoki Nakaya, Tomoya Hanibuchi, Shiho Amagasa, Hiroyuki Kikuchi, and Shigeru Inoue. Objective scoring of streetscape walkability related to leisure walking: Statistical modeling approach with semantic segmentation of google street view images. Health&Place, 66:...

    2020

  26. [34]

    Streetscore - predicting the perceived safety of one mil- lion streetscapes

    Nikhil Naik, Jade Philipoom, Ramesh Raskar, and Cesar Hi- dalgo. Streetscore - predicting the perceived safety of one mil- lion streetscapes. In CVPR, 2014. 1

    2014

  27. [35]

    The mapillary vistas dataset for semantic understanding of street scenes

    Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, 2017. 2

    2017

  28. [36]

    Counterfactual vqa: A cause- effect look at language bias

    Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xian- Sheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause- effect look at language bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12700–12710, 2021. 2

    2021

  29. [37]

    Evaluating the subjective percep- tions of streetscapes using street-view images

    Yoshiki Ogawa, Takuya Oki, Chenbo Zhao, Yoshihide Seki- moto, and Chihiro Shimizu. Evaluating the subjective percep- tions of streetscapes using street-view images. Landscape and Urban Planning, 247:105073, 2024. 1

    2024

  30. [38]

    Learning transferable visual models from natural language supervision

    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 NeurIPS, 2021. 1

    2021

  31. [39]

    Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer

    Ren ´e Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, 2020. 2, 3

    2020

  32. [40]

    Visual cot: Advancing multi-modal language models with a compre- hensive dataset and benchmark for chain-of-thought reason- ing

    Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuo- fan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Advancing multi-modal language models with a compre- hensive dataset and benchmark for chain-of-thought reason- ing. In Advances in Neural Information Processing ...

    2024

  33. [41]

    Is synthetic data all we need? benchmarking the robustness of models trained with synthetic images

    Krishnakant Singh, Thanush Navaratnam, Jannik Holmer, Si- mone Schaub-Meyer, and Stefan Roth. Is synthetic data all we need? benchmarking the robustness of models trained with synthetic images. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognit...

    2024

  34. [42]

    Svensson

    Marie K. Svensson. Sky view factor analysis – implications for urban air temperature differences. Meteorological Appli- cations, 11(3):201–211, 2004. 4

    2004

  35. [43]

    A new benchmark: On the util- ity of synthetic data with blender for bare supervised learn- ing and downstream domain adaptation

    Hui Tang and Kui Jia. A new benchmark: On the util- ity of synthetic data with blender for bare supervised learn- ing and downstream domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15954–15964, 2023. 5

    2023

  36. [44]

    Gemini 1.5: Unlocking mul- timodal understanding across millions of tokens of context

    Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking mul- timodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 4

    2024 arXiv

  37. [45]

    Negation: A pink elephant in the large language models’ room? arXiv preprint arXiv:2503.22395, 2025

    Tereza Vrabcov ´a, Marek Kadl ˇc´ık, Petr Sojka, Michal ˇStef´anik, and Michal Spiegel. Negation: A pink elephant in the large language models’ room? arXiv preprint arXiv:2503.22395, 2025. 4

    2025 arXiv

  38. [46]

    Al- varez

    Shihao Wang, Zhiding Yu, Xiaohui Jiang, Shiyi Lan, Min Shi, Nadine Chang, Jan Kautz, Ying Li, and Jose M. Al- varez. Omnidrive: A holistic vision-language dataset for au- tonomous driving with counterfactual reasoning. In Proceed- ings of the Computer Vision and Pattern Recogn...

    2025

  39. [47]

    Self-consistency improves chain of thought reasoning in lan- guage models

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in lan- guage models. In ICLR, 2023. 4

    2023

  40. [48]

    Chi, Quoc V

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V . Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In NIPS, 2022. 4

    2022

  41. [49]

    Alvarez, and Ping Luo

    Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. Segformer: Simple and effi- cient design for semantic segmentation with transformers. In NIPS, 2021. 3

    2021

  42. [50]

    Improve vision language model chain-of-thought reasoning, 2024

    Ruohong Zhang, Bowen Zhang, Yanghao Li, Haotian Zhang, Zhiqing Sun, Zhe Gan, Yinfei Yang, Ruoming Pang, and Yim- ing Yang. Improve vision language model chain-of-thought reasoning, 2024. 4

    2024

  43. [51]

    Negvqa: Can vision language models understand nega- tion? arXiv preprint arXiv:2505.22946, 2025

    Yuhui Zhang, Yuchang Su, Yiming Liu, and Serena Yeung- Levy. Negvqa: Can vision language models understand nega- tion? arXiv preprint arXiv:2505.22946, 2025. 2

    2025 arXiv

  44. [52]

    A study on the impact of visible green index and vegetation structures on brain wave change in residential landscape.Urban Forestry & Urban Greening, 64:127299, 2021

    Huaizhen Zhu, Fan Yang, Zhiyi Bao, and Xinge Nan. A study on the impact of visible green index and vegetation structures on brain wave change in residential landscape.Urban Forestry & Urban Greening, 64:127299, 2021. 4

    2021

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.