{"id":"b7d5e561-351f-40ca-a20f-58a5830eb45b","arxiv_id":"2412.18641","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ZenSVI provides an integrated, documented Python pipeline for acquiring, cleaning, analyzing, and visualizing street view imagery for urban science.","lead":"ZenSVI is a free, open-source Python package that unifies the entire street-view-image workflow, from downloading images from Mapillary, KartaView, and Amsterdam to computer-vision analysis and mapping. It aims to lower the technical barrier for urban researchers who want to use street-level imagery in their studies.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The case study's blur filter appears inverted: 'Laplacian variance over 100' removes sharp images, undermining the quality-assessment demonstration and the reproducibility claim unless clarified.","rationale":"The reader's weakest assumption centered on thresholds lacking sensitivity analysis and on treating missing values as suitable, which biases the demonstration. My concern overlaps on the Laplacian threshold but is more specific and potentially more serious: the direction of the threshold appears inverted relative to standard blur-detection practice. This does not overturn the reader's CONDITIONAL verdict, because the software's core contribution and public code remain valuable; however, it provides a concrete, checkable technical flaw that should be corrected before the case study is relied upon as evidence for the reproducibility and quality-assessment claims. I recommend keeping the verdict CONDITIONAL (i.e., no change from the reader), but the revision should address the blur-filter direction, not merely add sensitivity analysis. The proposed concrete test is inexpensive and directly settles whether the paper or the code is at fault.","tokens_in":27227,"tokens_out":3728,"duration_ms":36393,"concrete_test":"Inspect the ZenSVI source code on GitHub (likely in the computer vision or quality module) for the blur detection implementation, e.g., a function computing cv2.Laplacian(image, cv2.CV_64F).var(). Determine whether the suitability flag is set for images with variance below or above the threshold. Then run the blur detection on a small set of synthetically blurred and sharp images: if sharp images are labeled unsuitable (variance > 100 removed), the filter is inverted; if blurry images are removed (variance < 100), the paper has a typo. This single check settles whether the case study's quality filtering is correct.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that ZenSVI provides an integrated, reproducible SVI workflow, supported by a Singapore case study in Section 4. The load-bearing weakness is in the case study's quality filtering: Section 4.1 states 'Blurry images with Laplacian variance over 100 were removed.' In standard blur detection, the variance of the Laplacian is a focus measure: low variance indicates blur, high variance indicates sharp images. A threshold that removes images with variance over 100 would therefore discard the sharpest images and retain blurry ones, exactly contradicting the stated goal of removing unsuitable imagery and the paper's assertion that this 'filtering approach ensures a consistent and high-quality dataset.' If the published text is a typo and the code uses variance below 100 as the blur criterion, the paper's description is still wrong and the reproducibility claim is compromised because a user following the paper would implement the inverse filter. Either way, the demonstration of the integrated quality-assessment pipeline is not reliable as written, and the central claim of a reproducible end-to-end workflow is weakened until this discrepancy is resolved.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents ZenSVI, an open-source Python package that integrates the acquisition, processing, transformation, computer vision analysis, and visualization of street view imagery (SVI). The authors describe the package's modular architecture, compare it with existing urban-analysis and SVI tools, and demonstrate its functionality through a case study in Singapore that applies quality filtering and k-means clustering on crowdsourced Mapillary imagery.","tokens_in":27389,"tokens_out":4688,"duration_ms":38592,"significance":"If the software performs as described, ZenSVI fills a genuine gap in SVI research by providing a single, well-documented, open-source workflow that could lower technical barriers for researchers and improve reproducibility. The paper contributes a detailed comparison of existing tools, a clear description of the package's five sub-packages, extensive online documentation, and a concrete case study. A particular strength is the reuse of established computer vision models with transparent citations, which supports the reproducibility claim. However, the case study as written contains an apparent error in the blur-detection filter and lacks sensitivity analysis for key thresholds; these issues directly affect the credibility of the demonstration and, in turn, the paper's central claim of a reliable, reproducible end-to-end workflow.","major_comments":[{"comment":"The sharpness filter states that \"Blurry images with Laplacian variance over 100 were removed.\" In standard focus/quality assessment, the variance of the Laplacian is a focus measure: low variance indicates blur, while high variance indicates sharp imagery. As written, this filter would discard the sharpest images and retain blurry ones, contradicting the stated goal of removing unsuitable imagery and the claim that the filtering ensures a high-quality dataset. If this is a typographical error and the implemented test is variance below 100, the text must be corrected because a user following the paper would otherwise apply the inverse filter, compromising reproducibility. Either way, the case study's quality-assessment demonstration is unreliable as written.","section":"Section 4.1"},{"comment":"The quality thresholds (Laplacian variance, visual complexity score of 1.0, device speed at most 200 km/h) and the number of clusters (k=5) are set without sensitivity analysis. The paper mentions that the elbow method suggested a range of possible k values, but it does not report how the clustering results vary across that range, nor how the filtering outcomes change with the thresholds. Since the case study is the primary evidence for the pipeline's effectiveness, the absence of robustness analysis leaves the reliability of the integrated workflow unquantified.","section":"Sections 4.1-4.2"},{"comment":"The statement \"For each criterion, missing values (NA) were treated as suitable to avoid excessive data loss\" means that images with missing metadata or missing classification outputs are automatically retained. This treatment biases the suitability flag toward retaining incomplete records and could undermine the claim that the filtering approach \"ensures a consistent and high-quality dataset.\" The paper does not report how many images had missing values for each criterion, nor does it analyze the potential impact of this choice on the final 270,864-image dataset.","section":"Section 4.1"}],"minor_comments":[{"comment":"The row for the Mask2Former model reports four accuracy values with ambiguous associations: \"Achieves mean IoU of 84.5% (semantic) and 64.7% (panoptic) for semantic segmentation and 66.6% and 45.5% for panoptic segmentation.\" Please restructure the table to clearly map each metric (semantic vs. panoptic, Cityscapes vs. Mapillary Vistas) to its corresponding value.","section":"Table 5"},{"comment":"The claim that \"This study is the first to create such an integrated software not only in SVI research but also in urban imagery\" is strong and not fully substantiated by the comparison in Table 2, which lists several tools with overlapping (if narrower) functionality. The authors may wish to soften the claim or provide a more systematic basis for it.","section":"Section 1"},{"comment":"The description of the equisolid angle projection states that \"its pixel ratio directly corresponds to solid angles from a viewpoint,\" which is only exactly true under the ideal equisolid-angle model. The wording could be clarified to avoid overstating the geometric accuracy.","section":"Section 3.3"},{"comment":"The manuscript uses \"sub-package\" and \"subpackage\" inconsistently; please standardize the spelling throughout.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable software contribution for Computers, Environment and Urban Systems, and the self-citation pattern is consistent with the authors' established body of work on SVI. The main technical concern is the apparently inverted blur-detection threshold in the case study, which is fixable but currently undermines the reproducibility claim. I would recommend the editor request a major revision addressing that point and the related lack of sensitivity analysis for the case study parameters."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"ZenSVI is a genuinely useful integration: it wraps Mapillary, KartaView, and Amsterdam downloads, metadata analysis, image transformation, a stack of pre-trained CV models, and visualization into one documented Python package with tests and CI. That is the contribution, and it is real. The paper is a software paper, so I read it as a claim that this package exists, is usable, and standardizes a fragmented workflow. On that claim, it mostly delivers.\n\nThe strongest parts are the open code, the readable architecture, and the honest comparison table with existing single-purpose tools. The Singapore case study is a convincing end-to-end demo with 932k images. The authors also have a clear sense of the ecosystem they are building on, and the modular design gives the package a real shelf life.\n\nThe soft spots are mostly in the case study, and one of them is a genuine howler. The blur filter as written is backwards: 'Blurry images with Laplacian variance over 100 were removed.' Standard Laplacian variance is a focus measure; low variance means blur, high variance means sharp. So the text says they removed sharp images and kept blurry ones. If the code does the opposite, the text still needs a correction, and a user following the paper literally would implement the inverse filter. This is a reproducibility issue in the quality-control step, and it should be fixed before publication.\n\nThe other issues are the usual demo-grade choices: the Laplacian threshold, visual complexity minimum, speed cutoff, and k=5 are all reasonable but are presented without sensitivity analysis. Treating NA as 'suitable' in every filter will bias the retained set toward images with missing metadata, which is worth disclosing explicitly. The 'first integrated software' claim is a bit strong, but Table 2 does back it up as long as you read it as 'first to cover all five pipeline stages' rather than literally first ever.\n\nThe efficiency claims are not benchmarked, but that is a minor omission in a software paper of this type. The integrated models come from prior work, including the authors' own Global Streetscapes models; that is fine because the package is a wrapper, and the citations are appropriate.\n\nBottom line: this deserves peer review, and with a typo fix and a short robustness note it should be a solid publication. Anyone doing street view imagery work should know about ZenSVI, and I would cite it if I were building a similar pipeline.","headline":"A solid, genuinely integrative software paper whose case study contains a reversible but embarrassing blur-filter typo that should be fixed before publication.","tokens_in":27976,"tokens_out":1905,"would_cite":true,"duration_ms":19564,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"ZenSVI integrates the full street view imagery workflow into one open-source pipeline.","keywords":["Street-level imagery","Urban analytics","Python package","Computer vision","FAIR","Reproducibility","Crowdsourced imagery","Open-source software"],"falsifier":"Take a stratified sample of crowdsourced street view images from a new city, run ZenSVI's full pipeline, and compare its quality-suitability decisions, segmentation ratios, perception scores, and scene labels against manual annotations or reference measurements; material disagreement on held-out images would show that the integrated pipeline's reliability depends on the wrapper assumptions rather than being guaranteed by the package.","tokens_in":27014,"feed_emoji":"🏙️","tokens_out":6930,"duration_ms":58887,"temperature":0.7,"pith_summary":"ZenSVI is an open-source Python package that claims to integrate the whole street view imagery (SVI) workflow—downloading, metadata analysis, image transformation, computer vision feature extraction, visualization, and export—into one modular tool. The authors position it as the first integrated software of its kind for SVI and urban imagery research, arguing that current practice is fragmented into isolated, single-purpose scripts that hamper reproducibility and exclude researchers without strong programming skills. The paper demonstrates the package end-to-end on Singapore, downloading 932,596 crowdsourced images, filtering them to 270,864 daytime images of acceptable quality, extracting urban features, and clustering them into five distinct streetscape types. If the central claim holds, a researcher with minimal code could go from a place name to cleaned, analyzed, and mapped street-level indicators, which would make SVI studies more transparent and comparable across cities.","feed_headline":"One open package now runs street view analysis from download to map","feed_subtitle":"ZenSVI unifies acquisition, cleaning, computer vision, and mapping to make urban-image research reproducible.","key_machinery":"The load-bearing mechanism is the modular five-subpackage architecture with standardized inputs and outputs. Download wraps crowdsourced and city-provided APIs with parallel processing and checkpoints; Metadata Analysis augments platform metadata with derived indicators such as season, day/night status, relative camera angle, speed, and H3 hexagon identifiers at image, street, and grid levels; Transformation converts panoramas into perspective and fisheye views and generates point clouds from depth estimates; Computer Vision exposes pre-trained segmentation, classification, object detection, depth estimation, embedding, and low-level feature models behind a common interface; Visualization produces point, line, and hexagonal maps plus distribution plots. The case study chains these modules, using a Shannon-index visual complexity threshold, Laplacian sharpness, speed, lighting, view direction, and platform filters before k-means clustering with k=5.","core_discovery":"On the paper's own terms, the discovery is a design and implementation result: an integrated, free, and open-source software framework that covers the entire SVI analysis pipeline, where no such integrated tool previously existed for street-level or urban imagery. The package standardizes inputs and outputs across five sub-packages—download, metadata analysis, transformation, computer vision, and visualization—so that established computer vision models and platform APIs can be chained into a single reproducible workflow. The Singapore case study functions as a proof that the integrated pipeline works at scale: 932,596 images were downloaded, quality-filtered to 270,864 suitable daytime images, and clustered into five interpretable urban streetscape categories with mapped spatial patterns.","pith_inferences":["If ZenSVI becomes a widely used standard, the field may converge on common cleaning and feature-extraction protocols, making published SVI-derived indicators more directly comparable across studies; the paper points toward this potential but does not itself establish cross-city validity.","The wrapper-based integration inherits the biases and failure modes of the pre-trained models and platform APIs without revalidating them on target crowdsourced data, so a careful user should benchmark model outputs on a local sample before trusting downstream indicators.","The case study's thresholds (Laplacian variance above 100, visual complexity at least 1.0, speed at most 200 km/h, and k=5) are set without sensitivity analysis, and missing metadata is treated as suitable to avoid data loss; testing how cluster results and suitability flags shift with these choices would quantify the pipeline's reliability.","The same modular pipeline could extend beyond street-level imagery to satellite, aerial, and CCTV imagery, as the authors note as future work, but that would require new download and transformation modules and is not demonstrated here."],"forward_implications":["A researcher can reproduce a full SVI study from a single documented pipeline instead of stitching together ad hoc scripts, directly addressing the reproducibility gap the paper identifies.","Comparisons across cities and across crowdsourced versus commercial imagery become more feasible because data acquisition, cleaning, and feature extraction share one standardized interface.","Researchers without deep programming or deep-learning expertise can apply state-of-the-art segmentation, perception, scene classification, and object detection to street-level imagery, lowering the entry barrier the paper highlights.","The modular design means new imagery platforms, new computer vision models, and new visualization types can be added without rewriting the pipeline, so the tool can grow with the field.","Outputs saved as CSV and standard geospatial files allow downstream analysis in other software, so ZenSVI can act as a preprocessing and feature-extraction layer in larger workflows."],"supporting_citations":[{"why":"Documents the fragmentation and reproducibility gap in SVI research that ZenSVI is designed to fill.","marker":"(Biljecki and Ito, 2021)"},{"why":"Provides OSMnx, the one-stop open-source model and the street-network and place-name functionality ZenSVI builds on for downloading.","marker":"(Boeing, 2017)"},{"why":"Supplies the Mask2Former model used for semantic and panoptic segmentation in the computer vision sub-package.","marker":"(Cheng et al., 2022)"},{"why":"Supplies the Place Pulse 2.0 perception model that generates the six-dimension streetscape perception scores.","marker":"(Dubey et al., 2016)"},{"why":"Supplies the Places365 scene classification model used for scene categories and attributes.","marker":"(Zhou et al., 2017)"},{"why":"Provides the eight image-quality classifiers (weather, glare, time, panorama, platform, quality, reflection, view direction) used in filtering.","marker":"(Hou et al., 2024)"},{"why":"Supplies the GroundingDINO open-set object detector used to quantify urban elements.","marker":"(Liu et al., 2023)"},{"why":"Supplies the Depth Anything absolute depth estimator used for point cloud generation.","marker":"(Yang et al., 2024)"},{"why":"Provides the Urbanity network data tool and the visual complexity index concept adopted in the case study filtering.","marker":"(Yap et al., 2023)"}],"fun_headline_variants":["ZenSVI: one pipeline for street view analysis, from download to map","Open-source Python package unifies the entire street view analysis workflow","One codebase for street view imagery: acquisition, processing, and mapping","ZenSVI: end-to-end street view analysis made reproducible and scalable","From API to map: ZenSVI standardizes street view research pipelines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that wrapping established platform APIs and pre-trained computer vision models without revalidating them on the target crowdsourced images preserves their performance and therefore yields outputs trustworthy enough for downstream urban science.","fun_headline_variants_meta":{"raw":{"variants":["ZenSVI: one pipeline for street view analysis, from download to map","Open-source Python package unifies the entire street view analysis workflow","One codebase for street view imagery: acquisition, processing, and mapping","ZenSVI: end-to-end street view analysis made reproducible and scalable","From API to map: ZenSVI standardizes street view research pipelines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001212,"raw_usage":{"total_tokens":5015,"prompt_tokens":994,"completion_tokens":4021,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":3926}},"tokens_in":610,"tokens_out":4021,"duration_ms":25133,"temperature":1.0,"reasoning_tokens":3926,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:53:45.139188+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a stratified sample of crowdsourced street view images from a new city, run ZenSVI's full pipeline, and compare its quality-suitability decisions, segmentation ratios, perception scores, and scene labels against manual annotations or reference measurements; material disagreement on held-out images would show that the integrated pipeline's reliability depends on the wrapper assumptions rather than being guaranteed by the package.","supporting_citations":[],"review_version":1}