{"id":"db0a2b2e-38a4-455b-896d-958f743a5cd6","arxiv_id":"2607.20263","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Photo aggregation raises building-level Macro-F1 from 60.8% to 75.0%; adding POI context reaches 76.8%, confirming neighborhood data is a weak supplementary prior.","lead":"A building-inspection study in Qingdao finds that combining many photos of the same building lifts automated detection of housing problems much more than adding neighborhood business and amenity data, which provides only a small extra accuracy gain. The value is a scalable, building-level protocol for urban renewal that makes the limits of contextual data explicit.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"POI gain may be inflated by test-set selection of detector, radius, and features; the main multi-view aggregation result appears robust.","rationale":"The reader's weakest_assumption correctly identifies POI feature selection leakage, but the concern is broader: the entire model configuration (detector choice, POI radius, and feature screening) is selected using the same test folds on which performance is reported. This makes the +1.84-point gain an optimistic upper bound. The secondary inconsistency about which POI features were actually used further undermines the quantitative POI claim. Nevertheless, the multi-view aggregation result (DD 60.84% to MVVA 74.95%) is independent of POI features and the spatial Group CV protocol is appropriate for that comparison. The paper's main qualitative conclusion—that multi-view imagery drives the improvement and POI context is a modest, category-dependent supplement—is plausibly correct even if the specific +1.84 gain is not reliable. A conditional acceptance requiring nested model selection and clarification of the feature set is appropriate; the verdict should remain CONDITIONAL rather than moving to rejection.","tokens_in":28680,"tokens_out":10737,"duration_ms":89661,"concrete_test":"Implement a fully nested version of the MSCF experiment: within each fold of the Spatial Group CV, perform POI variable screening (§3.3.2) using only training-fold communities, and select the POI radius (and, if feasible, the detector) using only training-fold validation. Evaluate on the held-out fold and compare the out-of-fold MSCF vs MVVA Macro-F1 difference to the reported +1.84. Repeat over multiple random seeds to bound sampling noise. If the nested difference is ≤0 or not consistently positive, the reported POI gain is a selection artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reported MSCF-vs-MVVA gain of +1.84 points (Macro-F1 74.95% to 76.79%) is not the result of a pre-specified comparison. In §4.4 and Table 14, the authors select YOLOv8 + 1000m because that configuration achieved the highest MSCF F1 on the Spatial Group CV test folds—this is model selection on test data. Additionally, §3.3.2 screens POI variables using Pearson/Spearman correlations computed over all 92 communities, including those that later appear in test folds. Thus the +1.84 gain is the maximum over several choices (detector, POI radius, feature set), and the paper does not account for this multiplicity (Table 17 is explicitly descriptive). A secondary inconsistency compounds the concern: at 1000m, only 'Addition balconies windows' has significant POI correlations (Table 6), yet Table 15 shows MSCF gains for Pipeline damage, Illegal renovation, and Elevator addition—implying the implemented feature set differs from the per-category screening described in §3.3.2–§3.4. If the configuration had been fixed in advance or selected within each training fold, the POI contribution might vanish, undermining the second quantitative claim. The central multi-view aggregation result (60.84% to 74.95%) does not depend on POI selection and appears more robust.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a three-stage vision–POI fusion framework for building-level housing inspection in old residential communities, using 3,237 buildings and 25,608 field images across 92 communities in Qingdao. Image-level object detection is performed with one of several detectors (YOLOv8, YOLOv9c, YOLOv10s, YOLO11n, RT-DETR-L), and detections are aggregated into six building-level visual features plus a binary visual-existence indicator. POI-derived neighborhood context is constructed at 500 m, 1,000 m, and 1,500 m buffers, with Pearson/Spearman correlation screening under FDR correction. A cost-sensitive Random Forest fuses visual and POI features under community-isolated Spatial Group CV. The main reported results are: Multi-View Visual Aggregation raises building-level Macro-F1 from 60.84% (Direct Detection) to 74.95%; adding POI context (MSCF) raises Macro-F1 to 76.79%, a gain of +1.84 points. The paper interprets POI context as a modest, category-dependent, non-causal supplementary prior.","tokens_in":28985,"tokens_out":4293,"duration_ms":38646,"significance":"If the central results hold, the paper makes a useful applied contribution: it demonstrates that building-level inspection can be improved substantially by aggregating multi-view image detections, and it provides a careful, community-isolated evaluation protocol for the fusion classifier. The paper is also transparent in several respects: it releases code and data links, reports class-level confusion matrices for sparse categories (Table 16), and explicitly frames POI features as predictive context rather than causal determinants. The class-wise diagnostic tables and the honest reporting of the limited POI gain are strengths. However, the second quantitative claim (the +1.84 pp POI gain) rests on a feature-selection and model-selection protocol that is not fully nested within the spatial cross-validation folds, and there is a potential overlap between the object detector's training communities and the test communities of the downstream classifier. These issues affect the headline MSCF-vs-MVVA comparison and require correction before the POI-context claim can be accepted.","major_comments":[{"comment":"The POI contribution is quantified by comparing MSCF with MVVA under a configuration that is selected on the test folds. In §3.3.2, Pearson/Spearman correlations are computed over all 92 communities, including communities that later appear in test folds of the Spatial Group CV. In §4.4, the final setting 'YOLOv8 + 1000m' is chosen because it achieved the highest MSCF F1 on those same test folds (Table 14). The reported +1.84 pp gain is therefore the maximum over the POI feature set, buffer radius, and detector front-end, and Table 17 is explicitly descriptive. To support the claim that POI context provides an additional gain, the feature-selection step (including the FDR screening and the radius choice) must be nested inside each training fold, or the full configuration must be pre-registered before seeing test outcomes. Until then, the second quantitative claim is vulnerable to selectio","section":"§3.3.2, §4.4, Table 14, Table 17"},{"comment":"The Spatial Group CV described in §4.5 is applied to the Random Forest classifier, but the visual features are produced by object detectors trained once on the whole image dataset under an 8:2 split (§2.4, Table 10). If that 8:2 split was not performed at the same community level as the 92-community folds, or if detector training communities overlap RF test communities, then the MVVA/MSCF test features are not genuinely out-of-sample. The paper should clarify whether the detector training/validation split was spatially isolated, and if not, it should retrain the detector within each spatial fold or demonstrate that this overlap has a negligible effect. This concern affects both the MVVA and MSCF results, including the central multi-view aggregation claim.","section":"§2.4, §4.5, Table 10"},{"comment":"There is a mismatch between the POI screening results and the POI features used in MSCF. Table 6 reports that at the 1,000 m radius only 'Addition balconies windows' has significant POI correlations after FDR correction, yet Table 15 shows MSCF gains for Pipeline damage, Illegal renovation expansion, and Elevator addition under the main YOLOv8 + 1000m setting. The paper should state exactly which POI variables were used for each of the seven categories in the 1,000 m model. If variables that did not pass the FDR screening were nevertheless included, the screening criterion is not consistently applied; if no POI variables were included for those categories, the reported gains need a different explanation.","section":"§3.3.2, §3.4, Table 6, Table 15"}],"minor_comments":[{"comment":"The references 'Figure 4-3', 'Figure 4-4', and 'Table 4-6' appear to refer to Figure 7 and Table 14. These cross-reference labels should be updated to the actual figure and table numbers.","section":"§3.3.2, §4.4"},{"comment":"The phrase 'spatial plausibility test' is used for the correlation analysis; consider renaming it to 'association screening' to avoid implying that a correlation is a test of spatial plausibility.","section":"§4.4, §4.6.2"},{"comment":"The table reports descriptive paired differences without confidence intervals or significance tests. The accompanying caveat is welcome, but adding the standard deviation of the per-fold differences (or a paired bootstrap interval) would make the limited POI gain easier to interpret.","section":"Table 17"},{"comment":"The 8:2 'spatial heterogeneity ratio' for the detector training/validation split is described only briefly. Please specify whether the split was by community, building, or image, because this directly affects the interpretation of the downstream spatial CV results.","section":"§2.4, §4.1"}],"recommendation":"major_revision","confidential_remarks":"This is a work with a solid applied dataset and a generally careful evaluation of multi-view visual aggregation, but the POI-context claim is currently supported by a protocol that selects features, radius, and detector on the test folds. The paper's own Table 17 is honestly labeled descriptive, which limits the damage, but the main abstract claim of a +1.84 pp POI gain needs a clean nested validation. The detector-training overlap issue is also worth resolving, as it affects the stronger MVVA claim. I recommend major revision, not rejection: the issues are correctable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the main result survives scrutiny. Aggregating multi-view detections to building level lifts Macro-F1 from 60.84 to 74.95, and that gain does not depend on POI features or on any test-set model selection. The additional POI gain (+1.84 points at 1000m) is plausible but probably an overestimate, because two selection steps peek at the test folds.\n\nWhat's genuinely new: the HOUSED v2 building-linked dataset (3,237 buildings, 25,608 images), the explicit image-to-building aggregation protocol, and a mostly careful evaluation protocol — community-isolated spatial CV, class-level confusion matrices, and an honest discussion of sparse categories. I also give credit for reporting Table 17 as descriptive, not significance-tested, and for not overclaiming causality.\n\nNow the soft spots, in proportion. First, the POI feature screening in §3.3.2 uses Pearson/Spearman correlations over all 92 communities, including those that later appear in test folds. That is a leak that can inflate the MSCF-vs-MVVA gain. Second, the paper selects YOLOv8 + 1000m as the main setting because that combination had the highest test-fold F1 (Table 14) — again, model selection on test data. The +1.84 is effectively the max over detector/radius choices, so it's an upper bound, not an unbiased estimate. Third, there's an internal inconsistency: at 1000m only 'Addition balconies windows' has significant correlations, yet Table 15 reports MSCF gains for Pipeline damage, Illegal renovation, and Elevator addition. The implemented POI feature set seems to differ from the per-category screening described in §3.3.2–§3.4. That needs clarifying. Fourth, the Pipeline damage result rests on 23 positives; F1 of 21.43% is a screening outcome, not a recognition result, and the authors say as much.\n\nNone of this overturns the central conclusion. Multi-view aggregation is the load-bearing contribution; POI context is supplementary, category-dependent, and the paper already frames it that way. The fixes are straightforward: nest feature selection inside training folds, pre-specify or select the radius/detector within CV, and report confidence intervals for the POI gain.\n\nWho gets value: researchers working on building-level urban inspection, pedestrian-scale datasets, and spatial CV methodology. It deserves a serious referee round. I would send it to review with a request for major revision addressing the leakage and the screening inconsistency.","headline":"Multi-view aggregation is the real story; the POI gain is modest and likely inflated by selection on test data — fix the leakage and the paper holds up.","tokens_in":29487,"tokens_out":3487,"would_cite":false,"duration_ms":26740,"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":"The paper claims that automated building-level housing inspection works primarily by aggregating detections across multiple photographs of each building, with neighborhood POI context adding only a small, category-dependent correction.","keywords":["urban physical examination","residential building health","object detection","multi-view aggregation","points of interest","building-level inspection","spatial cross-validation","visual-GIS fusion"],"falsifier":"Re-run the full pipeline with POI feature selection (correlation screening and radius choice) performed inside each training fold only, then compare MSCF and MVVA under the same community-isolated folds. If the macro-F1 difference falls to zero or below, the claim that POI context adds predictive value fails for this dataset; a positive difference that survives nested selection would confirm it.","tokens_in":28562,"feed_emoji":"🏚️","tokens_out":7234,"duration_ms":58727,"temperature":0.7,"pith_summary":"Automated housing inspection at the building level can be substantially improved by aggregating object-detection outputs across the many photographs taken of each building, rather than by judging single images. The paper tests whether neighborhood context—the mix of shops, offices, and services around a community, derived from points-of-interest (POI) data—adds further predictive information. On a dataset of 92 old residential communities, 3,237 buildings, and 25,608 field images in Qingdao, multi-view aggregation raises building-level Macro-F1 from 60.84% to 74.95%. Adding POI context raises it to 76.79%, a modest, category-dependent gain; context alone reaches only 47.88%, so it cannot stand in for visual evidence. The practical claim is that POI context acts as a supplementary prior for prioritizing and interpreting inspection results, not as a causal or standalone signal.","feed_headline":"Multi-view photos, not POI context, drive housing inspection accuracy","feed_subtitle":"Multi-view photo aggregation does the heavy lifting; nearby POIs add only a small, category-dependent correction.","key_machinery":"The load-bearing object is the three-stage pipeline: (1) image-level object detection on each field photo; (2) building-level aggregation that converts per-image detections into a binary visual-existence flag plus six statistics (count, max/mean/median confidence, confidence standard deviation, and view-level occurrence rate); and (3) a post-correction classifier that fuses those visual features with POI-derived neighborhood features computed at 500m, 1000m, and 1500m buffers. The paper's main comparison isolates the contribution of each stage: Direct Detection (threshold any image), Multi-View Visual Aggregation (the aggregated visual features alone), Urban Functional Context Only (POI feat","core_discovery":"The discovery is that the main performance leap in building-level housing inspection comes from a simple aggregation step: take all field images linked to one building, run an object detector on each, and summarize the detections into a seven-component visual signature (whether anything was found, how many boxes, confidence statistics, and the fraction of views that contain the issue). This multi-view visual aggregation lifts Macro-F1 from 60.84% for per-image thresholding to 74.95%. Adding POI-derived neighborhood features through a cost-sensitive classifier under community-isolated spatial cross-validation yields 76.79%, an absolute gain of only +1.84 points over visual aggregation alone a","pith_inferences":["A testable extension: nest the POI feature-selection step inside each training fold of the spatial cross-validation. If the MSCF-minus-MVVA F1 gain disappears, the reported +1.84 points is partly a selection-leakage artifact; if it persists, the contextual prior is real.","The paper's emphasis on F1 understates what a probabilistic post-corrector could deliver; POI context may add more value through probability ranking (ROC/PR-AUC) than through binary label changes, and future work should report risk-ranking metrics separately.","Because POI gains are small, the framework's transfer value likely lies in the multi-view aggregation step; a second city with different planning and POI ecosystems could be used to test whether any POI radius remains beneficial.","Static POI counts cannot capture transient activity; incorporating mobile-signaling or foot-traffic data, which the paper names as future work, would test whether dynamic context beats static context before investing in larger image datasets."],"forward_implications":["A detector that wins on image-level mAP need not be the best front-end for building-level diagnosis; the paper finds the best image-level detector was not the best building-level front-end.","The 1000m POI buffer outperforms both 500m and 1500m, so neighborhood context has a scale sweet spot; the 1500m scale, despite the strongest univariate correlations, produces worse out-of-sample predictions.","Urban context alone is insufficient: POI-only input yields Macro-F1 of 47.88%, so POI should be used only as a supplementary feature on top of visual evidence.","Spatial leakage inflates results: ordinary cross-validation gives 81.50% Macro-F1 versus 76.79% under community-isolated cross-validation, so community-isolated evaluation is necessary for honest estimates.","For the sparsest categories, POI fusion improves screening precision but not reliable detection; pipeline damage reaches 21.43% F1 yet only 3 of 23 positive buildings are found."],"fun_headline_variants":["Multi-view photos, not POI, drive housing inspection gains","Building health checks: multi-view photos beat POI context","AI housing inspection: multi-view aggregation does heavy lifting","For building health, photo aggregation outweighs POI context","Housing inspection accuracy: multi-view photos lead, POI minor"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing assumption is that the POI feature selection—correlation screening across the 92 communities—does not use information from communities that end up in the held-out test folds of the spatial cross-validation; the paper does not nest feature selection inside folds, so the reported +1.84 point POI gain may be inflated.","fun_headline_variants_meta":{"raw":{"variants":["Multi-view photos, not POI, drive housing inspection gains","Building health checks: multi-view photos beat POI context","AI housing inspection: multi-view aggregation does heavy lifting","For building health, photo aggregation outweighs POI context","Housing inspection accuracy: multi-view photos lead, POI minor"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000213,"raw_usage":{"total_tokens":1296,"prompt_tokens":819,"completion_tokens":477,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":408}},"tokens_in":563,"tokens_out":477,"duration_ms":5071,"temperature":1.0,"reasoning_tokens":408,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T10:18:01.081957+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the full pipeline with POI feature selection (correlation screening and radius choice) performed inside each training fold only, then compare MSCF and MVVA under the same community-isolated folds. If the macro-F1 difference falls to zero or below, the claim that POI context adds predictive value fails for this dataset; a positive difference that survives nested selection would confirm it.","supporting_citations":[],"review_version":1}