Pith. sign in

REVIEW 5 major objections 5 minor 41 references

Interpretable Multimodal Framework for Human-Centered Street Assessment: Integrating Visual-Language Models for Perceptual Urban Diagnostics

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A vision-language street-assessment framework reports F1 = 0.84 on objective features and 89.3% agreement with resident perceptions, with a text rationale for each judgment.

desk verdict A promising but under-validated vision-language street-perception pipeline; the headline accuracy numbers do not follow from the described audit. read the letter →

arxiv 2506.05087 v1 pith:YYZTJNSR submitted 2025-06-05 cs.CV cs.CL

HaoTian Lan This is my paper · ORCID
classification cs.CVcs.CL
keywords multimodalurbananalyticshumanperceptionvision-languagealignmentinterpretableAIstreetscapeevaluationVisualGLM-6BGPT-4SDG11
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 tries to establish that a computer can judge streets the way residents do, and explain its judgment in plain language. The proposed Multimodal Street Evaluation Framework (MSEF) uses a large language model to convert aggregated resident survey answers into soft labels, then fine-tunes a smaller vision-language model to output both scalar scores and natural-language rationales. On 15,360 street-view images from Harbin, the authors report $F_1 = 0.84$ for objective features and 89.3% agreement with aggregated resident perceptions across socioeconomic strata. The payoff would be a scalable, interpretable urban diagnostic that also surfaces context-dependent trade-offs, such as informal commerce raising perceived vibrancy while lowering pedestrian comfort, that purely objective metrics miss.

What carries the argument

The load-bearing mechanism is the distillation loop from residents to labels to student model. GPT-4 translates aggregated Likert-scale survey answers and qualitative comments into soft scalar labels and explanatory phrases; those labels supervise the fine-tuning of VisualGLM-6B via rank-8 LoRA matrices and P-Tuning v2 prompt embeddings. A hybrid attention block then produces heat-maps and natural-language rationales aligned with audit protocols, giving each prediction a human-readable justification. This dual-output structure, scalar scores plus text rationales, is what lets the framework be both scalable and interpretable.

What would settle it

Collect street-view images from a city outside Harbin, gather independent resident Likert ratings for the same images, and compare MSEF's predictions to those ratings. If agreement drops to near chance while the objective-feature F1 stays high, the claim that the model generalizes human-centered perception to unseen neighborhoods collapses.

Watch

Extended reading notes

Core claim

The central claim is that the gap between objective street metrics and lived experience can be closed by a teacher-student distillation chain. GPT-4 reads street panoramas and resident questionnaire responses to produce soft labels; those labels fine-tune VisualGLM-6B through LoRA and P-Tuning v2. The resulting student model is reported to detect objective features with $F_1 = 0.84$ and to match aggregated resident ratings at 89.3% consistency in held-out, geographically stratified validation. The discovery goes beyond accuracy: the model reproduces nonlinear and contradictory perceptual patterns, including connectivity helping satisfaction only up to a saturation point, informal commerce boosting vibrancy while depressing comfort, and architectural transparency benefiting commercial streets but not residential ones, indicating that it reasons about land-use context instead of applying universal spatial heuristics.

Load-bearing premise

The result rests on the assumption that GPT-4's soft labels, distilled from aggregated resident survey answers, are a valid and sufficient ground truth for teaching the student model to perceive like residents; if those labels are biased or unrepresentative, the reported 89.3% agreement reflects the distillation, not genuine human-like perception.

Editorial extensions

If this is right

  • Municipal planning could run low-cost, city-wide perceptual audits from static street imagery, replacing or prioritizing expensive door-to-door surveys.
  • The dual-output design lets planners trace each score to a specific deficiency, such as narrow sidewalks or poor lighting, so interventions can be targeted rather than guessed at.
  • The inverted-U connectivity effect and the vibrancy-comfort trade-off imply that fixed thresholds for good urban design are unreliable, so guidelines should be location-sensitive.
  • Because only about 2% of the base model weights are updated, the fine-tuning approach is feasible for planning agencies with limited compute or data.

Reading between the lines

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

  • If the distillation logic holds, the same GPT-4-to-student schema could be reused for perceptual dimensions not covered here, such as noise annoyance, thermal comfort, or nighttime safety, given matching survey data.
  • The reported centralizing bias (predicted satisfaction interquartile range of about 0.9) suggests the model flattens extreme negative experiences toward neutral, so planners using it for equity allocation should treat low-score differences as conservative lower bounds.
  • The land-use-dependent effect of architectural transparency implies that visual heuristics learned in one city may not transfer zero-shot; cross-city deployment would likely require re-tuning the prompt embeddings with local labels.
  • A testable extension is to check whether the natural-language rationales are culturally portable, for example whether the same scene generates different rationales when GPT-4 is prompted to emulate residents from different cities.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes MSEF, a multimodal street evaluation framework that combines a vision-language model (VisualGLM-6B) with GPT-4 to output both objective feature scores and predicted resident satisfaction scores, accompanied by natural-language rationales. Training relies on GPT-4-generated soft labels synthesized from aggregated resident questionnaire responses, with VisualGLM-6B fine-tuned using LoRA and P-Tuning v2. The paper reports an F1 score of 0.84 for objective feature detection and 89.3% agreement with aggregated resident perceptions, and presents qualitative findings on nonlinear and context-dependent perceptual effects in a Harbin case study.

Significance. If the reported results were properly supported, MSEF would be a valuable contribution to scalable and interpretable urban perception modeling, combining physical feature extraction with resident-aligned subjective scoring and textual explanations. The use of parameter-efficient fine-tuning and attention-based rationales is a promising direction, and the focus on context-dependent perception is relevant to SDG 11. However, the current evaluation does not substantiate the central claims: the subjective validation appears circular, the objective F1 metric is not computable from the described audit, and numerous internal inconsistencies undermine reproducibility.

major comments (5)
  1. [Sections 3.1, 4.4, and 5.2] The subjective evaluation is circular. In Section 3.1, GPT-4 translates residents' questionnaire answers into soft labels that 'seed the fine-tuning of VisualGLM-6B.' Section 4.4 states that Q&A entries are synthesized to reflect the aggregated judgment of resident responses, using median ratings. Section 5.2 then evaluates the model against a 'consensus ground truth' formed by averaging multiple human participants' Likert ratings. Unless the participants in Section 5.2 are explicitly shown to be disjoint from the 320 residents of Section 4.3, and their ratings were excluded from the label-generation process, the 89.3% agreement measures how faithfully the student reproduces GPT-4's distillation of the survey, not whether the model aligns with residents. The paper does not provide such a statement of independence, so the central claim of reproducing aggregated resident perceptions is not established.
  2. [Sections 5.1 and 5.2] The claimed F1 score for objective features is not supported by the described audit. Section 5.1 explicitly states that 'Human reviewers did not label the images themselves but instead validated the model-generated scores by assessing whether the numeric outputs were reasonable.' F1 score requires ground-truth labels to compute precision and recall; a plausibility judgment does not constitute such labels. The 84% match rate and 92% fuzzy-check agreement are agreement measures, not classification metrics. To report F1=0.84, the authors must provide a labeled test set, define the classes, and present a confusion matrix.
  3. [Abstract, Sections 4.2, 5.2, 5.3, and 6] The manuscript contains major inconsistencies in reported dataset sizes and performance metrics that make the results irreproducible. The abstract and Section 5.2 state 'over 15,000 annotated street-view images' and a full dataset of 15,360 images; Section 4.2 reports 'over 3,500 images'; Section 5.3 reports 736 images; Section 6 evaluates on 143 panoramas. Similarly, the F1 score is 0.84 in the abstract, Section 5.2, and Section 7, but Section 3.1 reports an F1 of 0.89. Community counts also disagree: Section 5.1 uses 300 communities, while Section 4.1 says 30 communities were selected for fine-tuning. The authors must reconcile these numbers and specify exactly which dataset was used for each reported result.
  4. [Sections 5.2 and 6.1] The subjective agreement metric is ill-defined and likely inflated by score compression. The paper reports '89.3% consistency' but never defines this metric, and the Bland–Altman analysis is mentioned without reporting the actual limits of agreement, mean difference, or per-dimension statistics. Section 6.1 shows that predicted satisfaction scores are heavily compressed (median = 4.1, IQR = 0.9 on a 1–7 scale), which makes high agreement with human ratings easier to achieve, especially if the human ratings are also concentrated around the middle. The authors should report full Bland–Altman statistics, define the consistency measure, and compare against a trivial baseline that always predicts the training-set mean.
  5. [Sections 4.3, 5.1, and 6.1] The objective feature set is inconsistently defined, which affects the interpretability of the F1 claim. Section 4.3 lists seven physical attributes (sidewalk width, roadway width, greening level, degree of motorization, commercial activity density, sky openness, and presence of public facilities), while Section 6.1 states that MSEF produces a 9-dimensional objective feature vector (scaled 1–7). Section 5.2 refers to 'objective feature detection' without specifying which features or how detection is operationalized. The paper must clarify the exact number and definition of objective classes used for the reported F1 score.
minor comments (5)
  1. [References] Several references are duplicated, e.g., reference [8] is the same SAGAI paper as [6], and references [11]/[19] and [12]/[21] are duplicated. Please deduplicate the bibliography.
  2. [Section 3.1, Equations (1) and (2)] Equations (1) and (2) contain encoding artifacts (e.g., '??' and '', '') that obscure the mathematical definitions. Please ensure all equations render correctly.
  3. [Section 4.6] The paragraph beginning 'Resident-provided Likert-scale scores were normalized...' appears twice in Section 4.6; remove the duplicate.
  4. [Sections 3.1 and 6] The role of GPT-4 at inference time is unclear. Section 3.1 describes VisualGLM-6B as the lightweight inference 'student' and GPT-4 as the label generator, but Section 6 states that the framework 'integrates VisualGLM-6B and GPT-4 via LoRA and P-Tuning v2.' Clarify whether GPT-4 is used during deployment or only for training.
  5. [Sections 5.3 and 6] The relationship between the field deployment counts is confusing: Section 5.3 mentions 107 bidirectional observation points and 736 images, while Section 6 evaluates 143 panoramas from eight street segments. Please explain how these numbers relate and whether the deployment data are a subset of the evaluation data.

Circularity Check

1 steps flagged · score 6.0 of 10

The 89.3% subjective-agreement claim is an in-distribution distillation check: GPT-4 soft labels built from aggregated resident ratings seed VisualGLM-6B, and the same aggregated resident scores on the full dataset are then reported as independent agreement.

  1. fitted input called prediction [Sections 3.1, 4.4, and 5.2]
    "The resulting soft labels both offer an interpretable ground truth for auditors and seed the fine-tuning of VisualGLM-6B... we synthesized a single representative answer reflecting the aggregated judgment... the answer reflected the median rating... On the full dataset of 15,360 images, the framework achieved an F1-score of 0.84 in objective feature detection, while subjective ratings predicted by the model aligned with aggregated resident scores at 89.3% consistency."

    The supervision for VisualGLM-6B is generated by GPT-4 from aggregated resident Likert ratings, using the median rating and a contextual summary as the Q&A answer. The subjective audit then measures agreement with 'aggregated resident scores' on the 'full dataset of 15,360 images', which is the same aggregation operation and, as reported, the full training corpus. The 89.3% figure therefore reflects how faithfully the student reproduces GPT-4's distillation of the survey, not an independent test of whether visual features predict human perception.

full rationale

The core validation loop is circular as written. Section 4.4 converts resident Likert ratings into Q&A soft labels by having GPT-4 synthesize 'a single representative answer reflecting the aggregated judgment' and use the 'median rating.' Section 3.1 states these soft labels 'seed the fine-tuning of VisualGLM-6B.' Section 5.2 then reports 89.3% agreement between model predictions and 'aggregated resident scores' on the 'full dataset of 15,360 images.' Unless the Section 5.2 residents are a completely disjoint sample and the images are a held-out split, which the paper never states, this is an in-distribution evaluation against the very aggregation that produced the training target. The reported agreement therefore measures distillation fidelity to GPT-4's rendering of the survey, not independent prediction of human perception. The objective audit is similarly weakened by using reviewer plausibility rather than independent labels, though the subjective loop is the primary load-bearing circularity. Separate numerical inconsistencies (F1 0.89 in Section 3.1 vs 0.84 in Section 5.2; 3,500 vs 15,360 images) compound reproducibility concerns but are not circularity per se. No self-citation chain is load-bearing; the problem is the training-evaluation data loop.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the assumption that GPT-4-synthesized labels from resident surveys are valid supervision, and that a VisualGLM-6B student fine-tuned on them generalizes. No dataset or code is released, and no new physical or mathematical entities are introduced.

assumptions (4)
  • domain assumption Aggregated Likert responses from 320 residents represent the true street perception ground truth.
    Sections 4.3 and 4.4 use median and aggregate scores as supervision targets without participant-level reliability or inter-rater agreement analysis.
  • ad hoc to paper GPT-4-generated Q&A summaries preserve the semantic content of resident and expert judgments.
    Sections 4.4 and 4.5 state answers were synthesized with GPT-4, and these soft labels seed VisualGLM-6B fine-tuning in Section 3.1.
  • domain assumption Static daytime street-view imagery contains enough information to infer safety, comfort, and vibrancy.
    Sections 4.2 and 6.3 use only visual imagery; the paper itself notes temporal knowledge is missing, yet treats imagery as sufficient for the central claim.
  • domain assumption Community-level spatial holdout prevents image-level data leakage.
    Section 4.6 reserves 20% of communities, but field deployment images come from a new area with different distribution, and no leakage test is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Interpretable Multimodal Framework for Human-Centered Street Assessment: Integrating Visual-Language Models for Perceptual Urban Diagnostics." pith.science (2026). https://pith.science/paper/YYZTJNSR

@misc{pith2026250605087,
  author       = {Pith},
  title        = {Pith review of: Interpretable Multimodal Framework for Human-Centered Street Assessment: Integrating Visual-Language Models for Perceptual Urban Diagnostics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YYZTJNSR}},
  note         = {Machine review of arXiv:2506.05087}
}
read the original abstract

While objective street metrics derived from imagery or GIS have become standard in urban analytics, they remain insufficient to capture subjective perceptions essential to inclusive urban design. This study introduces a novel Multimodal Street Evaluation Framework (MSEF) that fuses a vision transformer (VisualGLM-6B) with a large language model (GPT-4), enabling interpretable dual-output assessment of streetscapes. Leveraging over 15,000 annotated street-view images from Harbin, China, we fine-tune the framework using LoRA and P-Tuning v2 for parameter-efficient adaptation. The model achieves an F1 score of 0.84 on objective features and 89.3 percent agreement with aggregated resident perceptions, validated across stratified socioeconomic geographies. Beyond classification accuracy, MSEF captures context-dependent contradictions: for instance, informal commerce boosts perceived vibrancy while simultaneously reducing pedestrian comfort. It also identifies nonlinear and semantically contingent patterns -- such as the divergent perceptual effects of architectural transparency across residential and commercial zones -- revealing the limits of universal spatial heuristics. By generating natural-language rationales grounded in attention mechanisms, the framework bridges sensory data with socio-affective inference, enabling transparent diagnostics aligned with SDG 11. This work offers both methodological innovation in urban perception modeling and practical utility for planning systems seeking to reconcile infrastructural precision with lived experience.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 36 canonical work pages

  1. [1]

    Subjective and objective measures of streetscape perceptions: Relationships with property value in Shanghai[J]

    Qiu W, Li W, Liu X, et al. Subjective and objective measures of streetscape perceptions: Relationships with property value in Shanghai[J]. Cities, 2023, 132: 104037

    work page 2023

  2. [2]

    Research on Urban Street Spatial Quality Based on Street View Image Segmentation[J]

    Gao L, Xiang X, Chen W, et al. Research on Urban Street Spatial Quality Based on Street View Image Segmentation[J]. Sustainability (2071-1050), 2024, 16(16)

    work page 2024

  3. [3]

    Evaluation of Pedestrian-Perceived Comfort on Urban Streets Using Multi-Source Data: A Case Study in Nanjing, China[J]

    Qin J, Feng Y, Sheng Y, et al. Evaluation of Pedestrian-Perceived Comfort on Urban Streets Using Multi-Source Data: A Case Study in Nanjing, China[J]. ISPRS International Journal of Geo-Information, 2025, 14(2): 63

    work page 2025

  4. [4]

    Measuring residents’ perceptions of city streets to inform better street planning through deep learning and space syntax[J]

    Wang L, Han X, He J, et al. Measuring residents’ perceptions of city streets to inform better street planning through deep learning and space syntax[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2022, 190: 215-230

    work page 2022

  5. [5]

    Built environment, urban vitality and social cohesion: Do vibrant neighborhoods foster strong communities?[J]

    Mouratidis K, Poortinga W. Built environment, urban vitality and social cohesion: Do vibrant neighborhoods foster strong communities?[J]. Landscape and Urban Planning, 2020, 204: 103951

    2020

  6. [7]

    Urban Safety Perception Through the Lens of Large Multimodal Models: A Persona-based Approach[J]

    Beneduce C, Lepri B, Luca M. Urban Safety Perception Through the Lens of Large Multimodal Models: A Persona-based Approach[J]. arXiv preprint arXiv:2503.00610, 2025

    arXiv 2025

  7. [8]

    Streetscape Analysis with Generative AI (SAGAI): Vision-Language Assessment and Mapping of Urban Scenes[J]

    Perez J, Fusco G. Streetscape Analysis with Generative AI (SAGAI): Vision-Language Assessment and Mapping of Urban Scenes[J]. Available at SSRN 5226191, 2025

    work page 2025

  8. [9]

    Quality of public space and sustainable development goals: analysis of nine urban projects in Spanish cities[J]

    Naya R B, de la Cal Nicolás P, Medina C D, et al. Quality of public space and sustainable development goals: analysis of nine urban projects in Spanish cities[J]. Frontiers of Architectural Research, 2023, 12(3): 477-495

    work page 2023

Show all 41 references
  1. [10]

    Assessing street-level urban greenery using Google Street View and a modified green view index[J]

    Li X, Zhang C, Li W, et al. Assessing street-level urban greenery using Google Street View and a modified green view index[J]. Urban Forestry & Urban Greening, 2015, 14(3): 675-685

    2015

  2. [11]

    Streetscore-predicting the perceived safety of one million streetscapes[C]//Proceedings of the IEEE conference on computer vision and pattern recognition workshops

    Naik N, Philipoom J, Raskar R, et al. Streetscore-predicting the perceived safety of one million streetscapes[C]//Proceedings of the IEEE conference on computer vision and pattern recognition workshops. 2014: 779-785

    2014

  3. [13]

    The digital life of walkable streets[C]//Proceedings of the 24th international conference on World Wide Web

    Quercia D, Aiello L M, Schifanella R, et al. The digital life of walkable streets[C]//Proceedings of the 24th international conference on World Wide Web. 2015: 875-884

    2015

  4. [15]

    Sustainability, Creativity, and Livability in Great East Asian Megacities: Insights from Twenty-First Century Tokyo, Seoul, and Chinese Creative Clusters[J]

    Pernice R. Sustainability, Creativity, and Livability in Great East Asian Megacities: Insights from Twenty-First Century Tokyo, Seoul, and Chinese Creative Clusters[J]. 2025

    2025

  5. [16]

    Gender, fear and the night-time city[M]//Fair Shared Cities

    Roberts M. Gender, fear and the night-time city[M]//Fair Shared Cities. Routledge, 2016: 49-64

    2016

  6. [18]

    The shortest path to happiness: Recommending beautiful, quiet, and happy routes in the city[C]//Proceedings of the 25th ACM conference on Hypertext and social media

    Quercia D, Schifanella R, Aiello L M. The shortest path to happiness: Recommending beautiful, quiet, and happy routes in the city[C]//Proceedings of the 25th ACM conference on Hypertext and social media. 2014: 116-125

    2014

  7. [19]

    Dubey A, Naik N, Parikh D, et al. Deep learning the city: Quantifying urban perception at a global scale[C]//Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part I 14. Springer International Publishing, 2016: 196-212

    2016

  8. [20]

    The city and the car[J]

    Sheller M, Urry J. The city and the car[J]. International journal of urban and regional research, 2000, 24(4): 737-757

    2000

  9. [21]

    Assessing the level of walkability for women using GIS and location-based open data: The case of new york city[J]

    Gorrini A, Presicce D, Choubassi R, et al. Assessing the level of walkability for women using GIS and location-based open data: The case of new york city[J]. Findings, 2021

    2021

  10. [22]

    How to study public life[M]

    Gehl J, Svarre B. How to study public life[M]. Washington, DC: Island press, 2013

    2013

  11. [23]

    Effects of skeletal streetscape design on perceived safety[J]

    Harvey C, Aultman-Hall L, Hurley S E, et al. Effects of skeletal streetscape design on perceived safety[J]. Landscape and Urban Planning, 2015, 142: 18-28

    2015

  12. [24]

    Dynamic Urban Imagery and Emotional Perception in Shanghai's North Bund: A Deep Learning Approach[J]

    Yao Z, Zhu T. Dynamic Urban Imagery and Emotional Perception in Shanghai's North Bund: A Deep Learning Approach[J]. Advances in Education, Humanities and Social Science Research, 2025, 13(1): 242-242

    2025

  13. [25]

    Exploring the coherence and divergence between the objective and subjective measurement of streetscape perceptions at the neighborhood level: A case study in Shanghai[J]

    Song Q, Fang Y, Li M, et al. Exploring the coherence and divergence between the objective and subjective measurement of streetscape perceptions at the neighborhood level: A case study in Shanghai[J]. Environment and Planning B: Urban Analytics and City Science, 2024: 23998083241292680

    2024

  14. [26]

    Standardized green view index and quantification of different metrics of urban green vegetation[J]

    Kumakoshi Y, Chan S Y, Koizumi H, et al. Standardized green view index and quantification of different metrics of urban green vegetation[J]. Sustainability, 2020, 12(18): 7434

    2020

  15. [27]

    Urban greenery and mental wellbeing in adults: Cross-sectional mediation analyses on multiple pathways across different greenery measures[J]

    Wang R, Helbich M, Yao Y, et al. Urban greenery and mental wellbeing in adults: Cross-sectional mediation analyses on multiple pathways across different greenery measures[J]. Environmental research, 2019, 176: 108535

    2019

  16. [28]

    Determinants and effects of perceived walkability: a literature review, conceptual model and research agenda[J]

    De Vos J, Lättman K, Van der Vlugt A L, et al. Determinants and effects of perceived walkability: a literature review, conceptual model and research agenda[J]. Transport reviews, 2023, 43(2): 303-324

    2023

  17. [29]

    How to Construct an Urban Color System? Taking the Historic Center of Macau as an Example[J]

    Zhang Y, Wang P, Wei W, et al. How to Construct an Urban Color System? Taking the Historic Center of Macau as an Example[J]. Buildings, 2024, 14(9): 2874

    2024

  18. [30]

    Learning transferable visual models from natural language supervision[C]//International conference on machine learning

    Radford A, Kim J W, Hallacy C, et al. Learning transferable visual models from natural language supervision[C]//International conference on machine learning. PmLR, 2021: 8748-8763

    2021

  19. [31]

    Urbanclip: Learning text-enhanced urban region profiling with contrastive language-image pretraining from the web[C]//Proceedings of the ACM Web Conference 2024

    Yan Y, Wen H, Zhong S, et al. Urbanclip: Learning text-enhanced urban region profiling with contrastive language-image pretraining from the web[C]//Proceedings of the ACM Web Conference 2024. 2024: 4006-4017

    2024

  20. [32]

    Uncertainties of human perception in visual image interpretation in complex urban environments[J]

    Kraff N J, Wurm M, Taubenböck H. Uncertainties of human perception in visual image interpretation in complex urban environments[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2020, 13: 4229-4241

    2020

  21. [33]

    Exploring the impact of urban built environment on public emotions based on social media data: a case study of Wuhan[J]

    Ma Y, Yang Y, Jiao H. Exploring the impact of urban built environment on public emotions based on social media data: a case study of Wuhan[J]. Land, 2021, 10(9): 986

    2021

  22. [34]

    Quantitative land price analysis via computer vision from street view images[J]

    Zhao C, Ogawa Y, Chen S, et al. Quantitative land price analysis via computer vision from street view images[J]. Engineering applications of artificial intelligence, 2023, 123: 106294

    2023

  23. [35]

    Evaluating the subjective perceptions of streetscapes using street-view images[J]

    Ogawa Y, Oki T, Zhao C, et al. Evaluating the subjective perceptions of streetscapes using street-view images[J]. Landscape and Urban Planning, 2024, 247: 105073

    2024

  24. [36]

    Timechat: A time-sensitive multimodal large language model for long video understanding[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Ren S, Yao L, Li S, et al. Timechat: A time-sensitive multimodal large language model for long video understanding[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024: 14313-14323

    2024

  25. [37]

    Urban Visual Appeal According to ChatGPT: Contrasting AI and Human Insights[J]

    Malekzadeh M, Willberg E, Torkko J, et al. Urban Visual Appeal According to ChatGPT: Contrasting AI and Human Insights[J]. arXiv preprint arXiv:2407.14268, 2024

    2024 arXiv

  26. [38]

    Multimodal Spatio-Temporal Data Visualization Technologies for Contemporary Urban Landscape Architecture: A Review and Prospect in the Context of Smart Cities[J]

    Han X, Li Z, Cao H, et al. Multimodal Spatio-Temporal Data Visualization Technologies for Contemporary Urban Landscape Architecture: A Review and Prospect in the Context of Smart Cities[J]. Land, 2025, 14(5): 1069

    2025

  27. [39]

    StreetviewLLM: Extracting Geographic Information Using a Chain-of-Thought Multimodal Large Language Model[J]

    Li Z, Xu J, Wang S, et al. StreetviewLLM: Extracting Geographic Information Using a Chain-of-Thought Multimodal Large Language Model[J]. arXiv preprint arXiv:2411.14476, 2024

    2024 arXiv

  28. [40]

    Omnigeo: Towards a multimodal large language models for geospatial artificial intelligence[J]

    Yuan L, Mo F, Huang K, et al. Omnigeo: Towards a multimodal large language models for geospatial artificial intelligence[J]. arXiv preprint arXiv:2503.16326, 2025

    2025 arXiv

  29. [41]

    LIME-Eval: Rethinking Low-light Image Enhancement Evaluation via Object Detection[J]

    Li M, Zhao H, Guo X. LIME-Eval: Rethinking Low-light Image Enhancement Evaluation via Object Detection[J]. arXiv preprint arXiv:2410.08810, 2024

    2024 arXiv

  30. [42]

    Urbangpt: Spatio-temporal large language models[C]//Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

    Li Z, Xia L, Tang J, et al. Urbangpt: Spatio-temporal large language models[C]//Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 2024: 5351-5362

    2024

  31. [43]

    UrbanMLLM: Joint Learning of Cross-view Imagery for Urban Understanding[J]

    Zhang X, Ouyang T, Shang Y, et al. UrbanMLLM: Joint Learning of Cross-view Imagery for Urban Understanding[J]

  32. [44]

    Revolutionizing urban safety perception assessments: Integrating multimodal large language models with street view images[J]

    Zhang J, Li Y, Fukuda T, et al. Revolutionizing urban safety perception assessments: Integrating multimodal large language models with street view images[J]. arXiv preprint arXiv:2407.19719, 2024

    2024 arXiv

  33. [45]

    Situation awareness in ai-based technologies and multimodal systems: Architectures, challenges and applications[J]

    Chen J, Seng K P, Smith J, et al. Situation awareness in ai-based technologies and multimodal systems: Architectures, challenges and applications[J]. IEEE Access, 2024

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.