{"id":"c74d932c-d8d1-4bac-b9c7-f2f4b3d0aea5","arxiv_id":"2412.11529","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The authors introduce the decentrality problem in cross-view geo-localization and present DReSS, a worldwide dataset with strong query-reference offsets, plus AuxGeo, an auxiliary-task method that achieves state-of-the-art results on DReSS, VIGOR, CVUSA, and CVACT.","lead":"This paper proposes DReSS, a large new dataset for matching street-view photos to satellite images when the query is offset from the satellite image center, a realistic setting the authors call 'decentrality.' It also introduces AuxGeo, a training method that uses bird's-eye-view and position-constraint auxiliary tasks to improve retrieval accuracy, outperforming prior methods on both the new dataset and existing benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DReSS's one-positive ground-truth assignment is under-specified: Table 2's pair counts vary with overlap and exceed the query count at 20%-50%, so the central decentrality comparison may rest on arbitrary labels.","rationale":"The paper has two intertwined contributions: a new dataset (DReSS) and a method (AuxGeo). The method-side story is reasonably well supported: the architecture is clear, the ablation in Table 9 shows that both BIM and PCM help, and the public-dataset numbers on CVUSA/CVACT/VIGOR are consistent with a strong baseline plus an auxiliary training signal. I do not see an internal inconsistency in the loss or the modules that would make the method unsound. The weak point is the dataset-side guarantee that DReSS actually measures what the paper claims. Section 3.1 defines the hit area and subsets only verbally; Section 3.4 gives capture details but not the positive-assignment rule; Section 3.5 assumes the labels already exist. Because DReSS is a new benchmark, its labels are the foundation for both the decentrality analysis and the 'outperforms previous methods on DReSS' claim. If the unique-positive assignment is arbitrary or non-reproducible, every DReSS number in Tables 5 and 6 is likewise non-reproducible. The Table 2 count discrepancy is the strongest concrete signal that the pipeline is under-specified: changing the overlap should not change the number of street-view queries, yet the reported pair counts exceed the query count at higher overlaps. This is exactly the kind of issue that is settled by releasing artifacts, not by argument. I therefore keep the reader's CONDITIONAL verdict: accept only if the authors release the metadata, the assignment script, and the subset definitions, and if the recounted labels reproduce the paper's tables. My concern is not that the authors are hiding a problem; it is that the current manuscript does not permit an independent check of the central DReSS claims, and one table already hints at an inconsistency.","tokens_in":15920,"tokens_out":12478,"duration_ms":119682,"concrete_test":"Release the DReSS metadata and label-generation script: panorama IDs with coordinates, tile footprints, the exact rule for choosing the unique positive tile, and the S1-S4 boundary thresholds. Independently run the script and verify that (i) every query receives exactly one positive, and (ii) the Table 1 and Table 2 subset counts reproduce exactly, including why the 20%-50% overlap counts sum to 175,170 instead of 174,934. Then recompute Table 5 and Table 6 using 'any tile whose footprint contains the query counts as a correct retrieval' as the positive set. If R@1 or the AuxGeo-vs-Sample4Geo margin changes by more than about 2 points on DReSS SAME, or if the per-subset improvement ordering in Table 6 changes, the paper's central comparison is not robust to the labeling choice.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing part of the central claim is the DReSS evaluation: AuxGeo 'mitigates the issue of large decentrality' and outperforms prior methods on DReSS. That evaluation requires each street-view query to have a well-defined 'best-matched' aerial reference and well-defined S1-S4 decentrality subsets. Section 3.1 defines the hit area conceptually and Section 3.5 defines the train/test splits, but neither specifies how the unique positive tile is chosen when a query point lies inside multiple 224x224 reference tiles at 12.5% overlap, nor how the S1-S4 boundaries are computed in geographic coordinates. The ambiguity is visible in Table 2: at 12.5% overlap the S1-S4 pair counts sum to exactly 174,934 (the number of street-view queries), but at 20%-50% overlap the same four counts sum to 175,170, implying duplicate or inconsistently counted positives. If the positive label is not a deterministic, one-to-one function of query and tile geometry, R@1 in Tables 5-6 is not a well-defined retrieval metric, and the reported performance gaps, especially the improvement pattern across S1-S4 in Table 6, may inherit an arbitrary labeling artifact. The gap between R@1 and Hit on DReSS (e.g., Sample4Geo SAME: 51.40 vs 55.89) confirms that top-1 retrievals which do cover the query are currently scored as errors; under a localization-based positive set the conclusions could differ. This concern is checkable and does not by itself invalidate the method, but it currently blocks verification of the central DReSS claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the 'decentrality' concept for cross-view geo-localization, defined as the offset between a street-view query and the center of its aerial reference image, and argues that existing datasets (CVUSA, CVACT, VIGOR) do not adequately cover large decentrality. It proposes DReSS, a new dataset of 174,934 street-view panoramas and 422,760 aerial images from eight cities, with a lower reference-tile overlap (12.5%) than VIGOR, and defines four decentrality subsets S1-S4. It also proposes AuxGeo, which adds two training-only auxiliary modules to a Sample4Geo baseline: a Bird's-eye view Intermediary Module (BIM) that uses BEV panoramas as an intermediate view, and a Position Constraint Module (PCM) that regresses query positions in aerial features at multiple scales. Experiments report SOTA or improved results on DReSS, VIGOR, CVUSA, and CVACT, with ablations showing contributions from both modules and no extra inference cost.","tokens_in":16301,"tokens_out":3339,"duration_ms":31569,"significance":"If the claims hold, the paper makes a useful contribution in two respects. First, DReSS would be the first public dataset specifically designed to stress-test cross-view geo-localization under large query-reference offsets, which is a realistic condition for pre-built reference databases. Second, AuxGeo demonstrates a credible recipe for using auxiliary signals (BEV transforms and position priors) at training time only, preserving inference efficiency. The paper's strengths include extensive comparisons across four datasets, ablations of both proposed modules, a generalization experiment across datasets, and the explicit promise of public code and data. The main caveat is that the central DReSS-based claims depend on a dataset construction protocol whose details are not yet fully specified.","major_comments":[{"comment":"The ground-truth assignment rule for DReSS is under-specified. Section 3.1 defines the hit area conceptually but does not state how a unique 'best-matched' aerial reference tile is chosen when a query point falls inside multiple overlapping 224x224 tiles, which occurs at all overlap levels below 100%. Table 2 shows that at 12.5% overlap the S1-S4 pair counts sum to exactly 174,934 (the number of street-view queries), but at 20%, 30%, 40%, and 50% overlap the same sums are 175,170, 175,172, 175,184, and 175,183 respectively, exceeding the query count. This implies that some queries are assigned to more than one positive tile, or that the counting procedure differs across overlap levels. Without a deterministic, one-to-one, geometrically defined assignment of positives, R@1 in Tables 5-6 is not a well-defined retrieval metric, and the per-subset decentrality analysis may inherit an arbitrary labeling artifact. Please specify the exact assignment rule (e.g., nearest tile center, fixed tie-breaking, or all covering tiles treated as positives) and provide the code or a precise algorithm used to generate the ground-truth pairs.","section":"Section 3.1, Table 2"},{"comment":"The boundaries of the four decentrality subsets S1-S4 and of the hit area are defined only visually. The paper does not give the geographic coordinates, distance intervals, or overlap-based formulas used to assign each pair to S1, S2, S3, or S4, nor does it define what 'with no interaction' means quantitatively. This makes the central result in Table 6 -- the progressively larger improvement of AuxGeo over Sample4Geo from S1 to S4 -- irreproducible. Please provide the exact set of inequalities (e.g., in meters or in fractions of tile size) that determine subset membership, and state whether the subset definitions are applied to the query position relative to the tile center, the tile corners, or the hit-area boundary.","section":"Section 3.1, Figure 2(a)"},{"comment":"The paper reports no error bars, multiple seeds, or significance tests for any of the main retrieval numbers. This is particularly consequential for the DReSS results: on VIGOR the improvement over Sample4Geo is 2.48 R@1 in SAME and 2.24 in CROSS, while on CVUSA the improvement is 0.12 R@1, and on DReSS SAME the improvement is 3.30 R@1. Without variance estimates, the claim that AuxGeo 'progressively' improves with increasing decentrality (Table 6) rests on a single run. Please report at least 3 seeds with mean and standard deviation for all tables where a comparison is made, or provide a statistical test (e.g., paired bootstrap over queries) for the headline DReSS and VIGOR comparisons.","section":"Tables 5-8 and Table 9"},{"comment":"The experimental protocol for the decentrality analysis is ambiguous. The text says 'Models were trained using data with various degrees of decentrality and evaluated across Subsets 1 to 4', but it does not state whether the models are trained on the full DReSS training set (which contains all decentrality levels) or on subsets restricted to particular decentrality ranges, nor whether the same model checkpoint is used for all four subset columns. If the former, the phrase 'trained using data with various degrees' is misleading; if the latter, the experimental setup needs to be described in full, including how the training subsets were balanced and whether the hyperparameters were re-tuned. Please clarify and, if multiple training protocols were used, report results for each.","section":"Section 5.2.2, Table 6"},{"comment":"The Hit metric on DReSS reveals a discrepancy that interacts with the ground-truth issue: for Sample4Geo SAME, R@1 is 51.40 while Hit is 55.89, and for AuxGeo SAME, R@1 is 54.70 while Hit is 59.20. This means a substantial number of top-1 retrievals are being scored as errors even though the retrieved aerial tile contains the query location, because they do not match the single 'best-matched' tile. If the evaluation instead used the standard localization-based positive set (all tiles covering the query), the relative ranking of methods and the per-subset improvement pattern could change. Please report results under both definitions (one-positive-per-query and all-covering-tiles) and discuss whether the conclusions of Tables 5 and 6 are robust to the choice of positive set.","section":"Section 3.5, Section 5.1.2"}],"minor_comments":[{"comment":"The sentence 'The panoramas are North-aligned, and each has a resolution of 2048 x 1024 pixels' should specify whether 'North-aligned' refers to the center of the panorama or the full 360-degree field of view, since this affects the BEV transformation and the orientation of the aerial images.","section":"Section 3.4"},{"comment":"Typo in the first sentence: 'We comapre our AuxGeo' should be 'We compare our AuxGeo'.","section":"Section 5.2"},{"comment":"The table lists 'Austrilia' for CVACT; this should be 'Australia'. Additionally, the 'Regions' row for DReSS says 'Urban and suburban', but Section 3.4 and the conclusion mention urban, suburban, and rural areas; please reconcile.","section":"Table 3"},{"comment":"The citation 'Zhang et al. (Zhang and Zhu, 2024)' appears in the text but the reference list contains 'Zhang, Q., Zhu, Y., 2024' with title 'Aligning geometric spatial layout...'; the in-text citation format is inconsistent with the reference entry and should be checked.","section":"Section 2.2.1"},{"comment":"Figure 5 is dense and the labels A1, A2, B are not explained in the caption. Please add a short caption description of the three panels, especially the data flow through BIM and PCM, to make the figure self-contained.","section":"Figure 5 and Section 4.2.1"},{"comment":"The loss notation uses the subscript/superscript convention _a^b inconsistently with the text: Equation (4) defines _street^aerial and _bev^street and _bev^aerial, but the text above refers to the InfoNCE loss in Equation (2) with different notation. Please standardize the subscript/ superscript convention.","section":"Section 4.2.4, Equation (4)"},{"comment":"The conclusion states that 'DReSS comprises cross-view images from eight global cities, representing diverse styles and encompassing extensive regions, including urban, suburban, and rural areas', but Table 3 lists only 'Urban and suburban'. Please align the descriptions.","section":"Section 7"}],"recommendation":"major_revision","confidential_remarks":"The paper's core contribution is a dataset, but the dataset construction details are not yet sufficient for independent verification. The inconsistencies in Table 2 (pair counts exceeding query counts at higher overlap) are the most concrete symptom. I would like the editor to ensure that the authors provide the actual ground-truth generation code or a precise algorithmic description, and that they address the positive-set ambiguity (one-tile vs. multiple-tile positives) before any final decision. The small margins on CVUSA/CVACT are not a concern by themselves, but the absence of variance estimates makes it hard to assess whether the headline improvements are robust. The paper does not appear to have any circularity problem: the method is evaluated on independent public benchmarks and the hyperparameters are not tuned to the test metrics."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nShort version: this paper makes a real contribution. DReSS is the first large cross-view benchmark that explicitly varies query-to-reference offset at scale, and AuxGeo is a simple, well-ablated training-only method that improves R@1 on four datasets with no extra inference cost. The decentrality concept is useful, and the global coverage is genuinely broader than VIGOR. I would take the DReSS claims seriously, but only after the authors document exactly how each street-view query gets its positive aerial tile.\n\nWhat is actually new: (1) DReSS, with 174,934 queries and 422,760 aerial images across eight cities, using 12.5% tile overlap and explicit S1–S4 decentrality subsets; (2) a clear demonstration that larger offsets change retrieval difficulty and database-size trade-offs; (3) AuxGeo, which adds an auxiliary BEV branch (BIM) and a multi-scale position-constraint head (PCM) to a standard contrastive backbone. The ablations in Table 9 are clean and show each module helps, with the combination best. Zero extra inference cost is a genuine plus.\n\nThe soft spot is the dataset construction. The paper defines \"best-matched\" and \"hit area\" conceptually, but never states how the unique positive tile is chosen when a query falls inside multiple overlapping tiles, which is unavoidable at 12.5% overlap. The stress-test note is correct: in Table 2, at 20–50% overlap the S1–S4 counts exceed the number of queries by a few hundred, so the counting is not one-positive-per-query, and the rule for the DReSS evaluation is ambiguous. If the assignment is arbitrary, the retrieval metric on DReSS is not well-defined. This does not sink the method, because the same assignment presumably applies to all compared methods, but it must be specified, and the dataset should be released with a deterministic rule and code. The S1–S4 boundaries are also not given in geographic coordinates.\n\nOther concerns are minor. There are no error bars or significance tests; the gap on CVUSA/CVACT is tiny (sub-1%), so a single seed could shift conclusions. The VIGOR gains are modest, while the DReSS-specific gains are larger and more convincing if the labeling holds.\n\nWho this is for: anyone in cross-view geo-localization or visual place recognition who cares about practical, non-centered matching. It is a benchmark paper first and a method paper second.\n\nRecommendation: send to peer review. The dataset and problem framing deserve referee time, and the method is solid enough. But the first decision letter should require a precise description of the positive-assignment and subset-boundary computation, plus code and data release, before the central claims can be verified.","headline":"Useful new DReSS benchmark and clean training-only AuxGeo method, but the dataset's ground-truth assignment is under-specified and must be documented before the central decentrality claims can be verified.","tokens_in":16848,"tokens_out":3519,"would_cite":true,"duration_ms":32362,"reading_group":"yes","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 cross-view geo-localization must handle the offset between a street-view query and the center of its aerial reference, and that its two training-only auxiliary modules—a bird's-eye view intermediary and a position…","keywords":["cross-view geo-localization","decentrality","street-view imagery","satellite imagery","image retrieval","bird's-eye view","position constraint","DReSS dataset"],"falsifier":"Re-derive the DReSS ground-truth matches from the panorama coordinates and the aerial tile grid, then recompute R@1 under that assignment; if the reported assignment cannot be reproduced, or the AuxGeo-versus-Sample4Geo gaps shrink or reverse under an alternative consistent assignment, the decentrality claims and the method's reported advantage would fail.","tokens_in":15732,"feed_emoji":"🛰️","tokens_out":7916,"duration_ms":60965,"temperature":0.7,"pith_summary":"The paper argues that real-world cross-view geo-localization must handle decentrality, the offset between where a street-view photo was taken and the center of the aerial image used as its reference, and that existing datasets mostly ignore it. To make the problem measurable, the authors build DReSS, an eight-city dataset of 174,934 street panoramas and 422,760 satellite tiles with only 12.5% overlap between reference tiles, so queries frequently sit far from tile centers. They then propose AuxGeo, which trains the network with two auxiliary tasks: a bird's-eye view intermediary that turns the hard 'cross-view plus offset' problem into two simpler ones, and a position constraint module that supervises the predicted location of the query on the aerial image. The paper reports that AuxGeo beats prior methods on DReSS, with larger gains at higher decentrality, and reaches state-of-the-art results on CVUSA, CVACT, and VIGOR, while adding no extra computation at inference.","feed_headline":"AuxGeo boosts off-center street-view localization by 3-4 points","feed_subtitle":"New 8-city dataset exposes the query-offset problem; a BEV intermediary and position constraints lift retrieval accuracy.","key_machinery":"The two load-bearing mechanisms are the Bird's-eye view Intermediary Module (BIM) and the Position Constraint Module (PCM). BIM converts each street-view panorama into a bird's-eye view image by a geometry-only transform and adds contrastive losses between street-BEV and aerial-BEV pairs, splitting 'cross-view with decentrality' into a cross-view subproblem and a decentrality subproblem. PCM uses the known geo-tag offset as supervision: it computes similarity maps between coarse street/BEV features and multi-scale aerial features, softly regresses the query's position on the aerial image at three pyramid levels, and applies L2 losses. Both modules are training-only; at inference the model is the plain backbone plus global pooling.","core_discovery":"On the paper's own terms, the central discovery is that decentrality is a separable, measurable obstacle: retrieval accuracy collapses as query-to-center offset grows, and a training-only combination of a BEV intermediary and multi-grained position supervision recovers a meaningful part of that loss. The evidence is the DReSS dataset, where reference tiles overlap at 12.5% and the hit area is subdivided into four decentrality subsets; on those subsets Sample4Geo's R@1 falls from 72.74 (S1) to 35.49 (S4), while AuxGeo improves over it by 1.68, 1.76, 3.62, and 4.01 points respectively. AuxGeo also outperforms the same baseline on VIGOR, CVUSA, and CVACT, so the auxiliary tasks help in both offset and center-aligned regimes.","pith_inferences":["If the DReSS query-to-tile assignment is reproducible from public geo-tags, the dataset could become a standard stress test, letting future methods report accuracy per offset bin rather than one average.","The training-only use of geometric priors suggests a general recipe: spatial supervision can be injected through auxiliary branches and discarded at inference, so a natural extension is applying BIM/PCM to drone-to-satellite or ground-to-map retrieval.","Because PCM already regresses the query's position within the aerial tile, the same network could return a continuous location estimate alongside the retrieved tile, turning retrieval into coarse localization at no extra inference cost.","The paper's efficiency argument implies practitioners could deliberately choose a low-overlap reference database and pay the accuracy cost only where the model cannot handle it; the paper does not provide a full Pareto curve of accuracy versus database size beyond the five overlap settings."],"forward_implications":["On DReSS, AuxGeo raises R@1 over Sample4Geo by 3.30 points in SAME and 4.21 points in CROSS, and the per-subset gains grow with decentrality (1.68, 1.76, 3.62, and 4.01 points across S1-S4).","AuxGeo also reaches the best reported results on VIGOR, CVUSA, and CVACT, so the auxiliary tasks do not trade away performance in center-aligned or modest-offset regimes.","Because BIM and PCM are training-only, deployment of AuxGeo costs the same as the baseline backbone: one forward pass with global average pooling.","Constructing reference databases at 12.5% overlap instead of 50% cuts the reference set by roughly a factor of 3.4, and the paper shows that methods must be decentrality-aware to benefit from that efficiency.","The paper's own limitation is that accuracy under extremely high decentrality remains low, so decentrality-aware retrieval is improved but not solved."],"supporting_citations":[{"why":"Supplies the VIGOR benchmark and the SAME/CROSS protocol that DReSS extends; its hit-area definition is the baseline for decentrality.","marker":"(Zhu et al., 2021)"},{"why":"Provides Sample4Geo, the training backbone and hard-negative baseline that AuxGeo builds on and outperforms on every dataset.","marker":"(Deuser et al., 2023)"},{"why":"Supplies the geometry-only transform that turns street panoramas into BEV images for the BIM.","marker":"(Wang et al., 2024)"},{"why":"Introduced the polar-transform family of methods that fail under decentrality, motivating the new approach.","marker":"(Shi et al., 2019)"},{"why":"FRGeo feature recombination is the strongest recent comparison and is shown to lose accuracy on large-decentrality settings.","marker":"(Zhang and Zhu, 2024)"},{"why":"CVUSA provides one of the center-aligned public datasets used to show AuxGeo retains state-of-the-art performance.","marker":"(Zhai et al., 2017)"},{"why":"CVACT provides the second center-aligned benchmark and city-scale test set used for evaluation.","marker":"(Liu and Li, 2019)"}],"fun_headline_variants":["DReSS dataset exposes decentrality problem in geo-localization","AuxGeo's BEV intermediary recovers off-center accuracy by 3-4 pts","Off-center street-view matching improved via multi-metric optimization","Position constraint module boosts cross-view retrieval under offset"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The DReSS evaluation depends on assigning each street-view query to exactly one 'best-matched' aerial reference tile, but the paper does not specify how that assignment is made when multiple overlapping tiles cover a query location, so a noisy or biased assignment would distort every reported accuracy and method comparison.","fun_headline_variants_meta":{"raw":{"variants":["DReSS dataset exposes decentrality problem in geo-localization","AuxGeo's BEV intermediary recovers off-center accuracy by 3-4 pts","Off-center street-view matching improved via multi-metric optimization","Position constraint module boosts cross-view retrieval under offset"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000569,"raw_usage":{"total_tokens":2734,"prompt_tokens":1027,"completion_tokens":1707,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":643,"completion_tokens_details":{"reasoning_tokens":1643}},"tokens_in":643,"tokens_out":1707,"duration_ms":13228,"temperature":1.0,"reasoning_tokens":1643,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:49:58.546992+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-derive the DReSS ground-truth matches from the panorama coordinates and the aerial tile grid, then recompute R@1 under that assignment; if the reported assignment cannot be reproduced, or the AuxGeo-versus-Sample4Geo gaps shrink or reverse under an alternative consistent assignment, the decentrality claims and the method's reported advantage would fail.","supporting_citations":[{"cited_title":", author Yang, T","cited_arxiv_id":null,"evidence_quote":"Supplies the VIGOR benchmark and the SAME/CROSS protocol that DReSS extends; its hit-area definition is the baseline for decentrality."},{"cited_title":", author Habel, K","cited_arxiv_id":null,"evidence_quote":"Provides Sample4Geo, the training backbone and hard-negative baseline that AuxGeo builds on and outperforms on every dataset."},{"cited_title":", author Xu, R","cited_arxiv_id":null,"evidence_quote":"Supplies the geometry-only transform that turns street panoramas into BEV images for the BIM."},{"cited_title":", author Liu, L","cited_arxiv_id":null,"evidence_quote":"Introduced the polar-transform family of methods that fail under decentrality, motivating the new approach."},{"cited_title":", author Zhu, Y","cited_arxiv_id":null,"evidence_quote":"FRGeo feature recombination is the strongest recent comparison and is shown to lose accuracy on large-decentrality settings."},{"cited_title":", author Bessinger, Z","cited_arxiv_id":null,"evidence_quote":"CVUSA provides one of the center-aligned public datasets used to show AuxGeo retains state-of-the-art performance."},{"cited_title":", author Li, H","cited_arxiv_id":null,"evidence_quote":"CVACT provides the second center-aligned benchmark and city-scale test set used for evaluation."}],"review_version":1}