{"id":"b98714c7-7451-49a1-9ddf-02fd2eaa6038","arxiv_id":"2608.11093","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A literature survey and benchmark of cross-view feature matching methods, organized by a new taxonomy and evaluated under partially consistent protocols.","lead":"This survey organizes the cross-view feature matching literature into a taxonomy and benchmarks representative sparse, semi-dense, and dense methods on standard datasets. It finds that dense methods are most accurate but costly, sparse methods are detector-limited, and semi-dense methods often generalize best to localization tasks.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ScanNet pose benchmark confounds training domain with method family: dense methods are ScanNet-trained while sparse/semi-dense are not, so the claimed accuracy hierarchy may be a domain-shift artifact.","rationale":"The strongest claim is the benchmarking-derived hierarchy, and the load-bearing assumption is that the comparison is fair across method families. The reader's weakest assumption captured exactly this; I agree and sharpen it to a specific confound: Table I's own note reveals that dense methods are ScanNet-trained while sparse and most semi-dense methods are not, making the ScanNet comparison an uncontrolled test of training domain rather than algorithmic paradigm. The resulting gap (RoMa v2 73.8 vs LightGlue 51.8 AUC@20) is large and could easily be driven by in-domain training. The underlined semi-dense entries provide internal evidence that ScanNet training itself raises scores by roughly 10 points, which is about the difference separating many paradigm classes. The survey's taxonomy and qualitative review remain sound, but the empirical hierarchy claim needs to be hedged or backed by matched-training re-evaluation. Since the reader already issued a conditional verdict, my analysis reinforces that judgment without changing it.","tokens_in":35603,"tokens_out":4528,"duration_ms":42935,"concrete_test":"Run a controlled ScanNet experiment: train LightGlue (SuperPoint backbone, same hyperparameters as MegaDepth) on the ScanNet training split, then evaluate on the same ScanNet test pairs as Table I. Compare its AUC@20 to the MegaDepth-trained LightGlue (51.8) and to RoMa v2 (73.8). If the ScanNet-trained sparse model gains >=10 AUC points, the dense-vs-sparse gap is largely a training-domain artifact. Equivalent test: evaluate a MegaDepth-only RoMa checkpoint on ScanNet and check whether its AUC drops toward semi-dense levels.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central empirical claim (Section III.C, Table I) is that a consistent hierarchy holds: dense > semi-dense > sparse for relative pose accuracy. But the table's own note admits dense methods are trained separately on each dataset, while sparse and most semi-dense methods are evaluated with a single MegaDepth-trained model; several semi-dense ScanNet entries are also from ScanNet-trained models (underlined). On ScanNet, the comparison therefore does not control for training data. RoMa v2 reaches AUC@20=73.8 vs LightGlue's 51.8; the magnitude of this gap is typical of domain shift (ScanNet is indoor, MegaDepth outdoor). The paper nonetheless concludes 'dense methods consistently outperform semi-dense approaches, which in turn generally surpass sparse methods' and uses this to motivate the paradigm trade-off. Since in-domain training, not matching density, could explain most of the ScanNet advantage, the hierarchy is not established on the very benchmark where it is largest. The same confound affects the underlined semi-dense entries: EcoMatcher (ScanNet-trained) at 64.0 vs ELoFTR (MegaDepth-trained) at 53.6 suggests training data moves scores by ~10 AUC points.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper is a survey of cross-view feature matching that proposes a six-part taxonomy (feature extraction, sparse/semi-dense/dense matchers, multi-type feature matchers, vision-foundation-model methods, training strategies, and robust estimation), reviews recent methods within that taxonomy, and presents a benchmark on MegaDepth, ScanNet, HPatches, Aachen Day-Night, and InLoc. The authors claim that the benchmark reveals a consistent performance hierarchy: dense methods achieve the highest pose and homography accuracy at higher computational cost, sparse methods are efficient but detector-limited, and semi-dense methods perform best overall in visual localization.","tokens_in":35844,"tokens_out":7697,"duration_ms":72371,"significance":"The survey's taxonomic organization is useful and its literature coverage is broad; the review of learning paradigms, VFM-based methods, and robust estimation provides a reasonable entry point for the field. The authors are also transparent in Table I about some training and code-availability caveats, which is a strength. However, the central empirical claims are currently not established because the benchmark does not control for training domain and dataset version across the compared methods. If the authors either add controlled comparisons or substantially weaken the hierarchy claims, the survey would be a valuable reference. The paper contains no derivations or machine-checked proofs; its value rests on the accuracy of the taxonomy and the credibility of the benchmark.","major_comments":[{"comment":"The claimed dense > semi-dense > sparse hierarchy for relative pose is confounded by training domain on ScanNet. The table's note states that dense models are trained separately on each dataset, while sparse and most semi-dense methods use a single MegaDepth-trained model, and several semi-dense ScanNet results (underlined) come from ScanNet-trained models. RoMa v2's AUC@20 of 73.8 versus LightGlue's 51.8 is therefore not a controlled comparison of matching density; the same table shows EcoMatcher (ScanNet-trained) at 64.0 versus ELoFTR (MegaDepth-trained) at 53.6, indicating that training data alone can move AUC by roughly 10 points. Please add ScanNet-trained sparse/semi-dense baselines or MegaDepth-trained dense evaluations, or explicitly restrict the hierarchy claim to settings with matched training data.","section":"Section III.C, Table I"},{"comment":"The visual localization comparison mixes dataset versions: the settings paragraph says sparse methods use Aachen Day-Night v1.0 while semi-dense and dense methods use v1.1, yet Table III reports a single Aachen Day-Night column group. Because v1.0 and v1.1 have different reference sets and query sets, the sparse versus semi-dense/dense comparisons on this benchmark are not mutually comparable. Please use one dataset version for all methods, or report results separately by version and avoid cross-category claims when versions differ.","section":"Section III.E, Table III"},{"comment":"The abstract and Section III describe a 'unified experimental benchmarking' under 'consistent protocols', but the Table I note admits that several semi-dense rows are literature-reported values from non-public code with unknown parameter counts, and dense methods keep their original image settings. This means the benchmark is not a single controlled implementation. Please state explicitly which rows were re-run by the authors, provide evaluation code and logs, and replace 'unified' and 'consistent' with a more precise description that lists the harmonized settings and the admitted exceptions.","section":"Section III, Table I note"}],"minor_comments":[{"comment":"The statement that dense methods 'attain the highest overall performance' is too strong: IMAmatch, a semi-dense method, achieves AUC@3px=72.4 versus RoMa's 72.2 and is within 0.2 AUC at the other thresholds. Please soften the wording to 'comparable or slightly higher'.","section":"Section III.D, Table II"},{"comment":"The underlining used to mark ScanNet-trained semi-dense results is not visible in the provided text; ensure the final PDF renders the underlining clearly and add an explicit legend in the table caption.","section":"Section III.C, Table I"},{"comment":"Parameter counts are missing for SEM and EcoMatcher; the efficiency discussion in Section III.C should avoid drawing parameter-cost conclusions from incomplete data.","section":"Section III.C, Table I"},{"comment":"The taxonomy labels CasMTR and IMAmatch as semi-dense methods that bridge to dense matching, while DKM and RoMa are classified as dense; please state the classification criterion explicitly so that coarse-to-fine architecture alone does not become an ambiguous discriminator.","section":"Section II"},{"comment":"The evaluation uses Aachen Day-Night v1.0 for sparse methods and v1.1 for semi-dense and dense methods; even if the final version reports each separately, the table should include a version column so readers can see the protocol difference immediately.","section":"Section III.E"}],"recommendation":"major_revision","confidential_remarks":"The benchmark and the survey include a substantial number of works from the authors' own groups (e.g., NCTR, ParaFormer, JamMa, SemaGlue, RCM, CoMatch, and others), and the selection of 'representative' methods is not justified. I am not questioning the authors' integrity, but the editor may wish to ask for a disclosure of selection criteria and of the authors' methodological overlap with the compared entries. Also, several references are 2026 arXiv preprints; please verify that these are publicly available and correctly cited before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is a genuinely useful survey of cross-view feature matching, and the taxonomy is the most structured I've seen for this area, especially the split of VFM-based methods into DINO, SAM, diffusion, and geometry-VFM categories. The benchmarking effort, covering MegaDepth, ScanNet, HPatches, Aachen, and InLoc, gives a practical snapshot of the field. If you need a reference for the current landscape, this will serve.\n\nThat said, the central empirical claim — that dense methods consistently outperform semi-dense, which outperform sparse — is not established on ScanNet. The table note admits it: dense methods are trained separately on each dataset, while sparse and most semi-dense methods use a single MegaDepth-trained model. RoMa v2's 73.8 vs LightGlue's 51.8 on ScanNet AUC@20 is exactly the kind of gap you'd expect from training-data domain shift, not necessarily from matching density. The paper underlines some ScanNet-trained semi-dense entries, so the authors know the issue, but the conclusion still states the hierarchy as if it were controlled. The HPatches and Aachen results are less affected, and those do support a milder version of the hierarchy.\n\nAlso, parameter counts are missing for several methods, the selection of representative methods isn't justified, and no benchmark code or repository is released. These are fixable, but they undercut the \"fair and comprehensive\" claim.\n\nThe survey itself is solid and worth sending to review, but the benchmarking section needs a major revision: either compare methods under matched training conditions, or explicitly frame the ScanNet table as a cross-domain comparison and keep the hierarchy claim for the controlled parts. If the authors soften the language and release the evaluation code, this becomes a reference-level survey.\n\nI'd bring it to a reading group to discuss the confound, and I'd cite it as a field overview. Deserves a serious referee, but not in its current form.","headline":"Useful survey and taxonomy, but the ScanNet benchmark confounds training domain with method family, so the central dense-vs-sparse hierarchy claim is overstated.","tokens_in":36303,"tokens_out":2308,"would_cite":true,"duration_ms":20925,"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":"A unified benchmark of image matchers finds accuracy rises with matching density, while semi-dense methods win at localization.","keywords":["cross-view feature matching","local feature matching","vision foundation models","sparse matcher","semi-dense matcher","dense matcher","visual localization","robust estimation"],"falsifier":"Retrain every sparse and semi-dense method on ScanNet's own training split, the same data dense methods see, and re-run relative pose estimation. If the dense advantage at AUC@20° shrinks below the gap reported in Table I, or if a sparse method overtakes a dense one, the claimed paradigm hierarchy reflects training conditions rather than matching density.","tokens_in":35456,"feed_emoji":"🎯","tokens_out":5471,"duration_ms":46182,"temperature":0.7,"pith_summary":"This survey tries to give the field of cross-view feature matching—finding corresponding points across images taken from very different viewpoints—one shared map. It proposes a six-axis taxonomy covering feature extraction, single-type matchers, multi-type matchers, vision-foundation-model methods, training strategy, and robust estimation, then runs representative methods through a unified benchmarking protocol on standard datasets. Its central finding is a consistent performance hierarchy: dense methods reach the highest accuracy but cost the most, sparse methods stay efficient yet are capped by keypoint detector quality, and semi-dense methods give the best overall visual-localization performance. The paper matters because it translates a fragmented literature into a single ordering that researchers can use to choose methods and target gaps.","feed_headline":"Dense matchers win on accuracy; semi-dense on localization","feed_subtitle":"A unified benchmark of sparse, semi-dense, and dense feature matchers shows detector quality caps sparse methods.","key_machinery":"The load-bearing object is the six-axis taxonomy, which treats cross-view feature matching not as a single algorithm family but as a layered design space spanning feature extraction, sparse/semi-dense/dense matching, multi-type feature fusion, vision-foundation-model use, training strategy, and robust estimation. The matching-paradigm split (sparse, semi-dense, dense) is the axis that does the explanatory work in the benchmark. The second mechanism is the unified evaluation protocol: one detector for sparse methods, one resizing rule for semi-dense methods, per-paper settings for dense methods, shared RANSAC thresholds, and the standard hierarchical localization pipeline, which together allow the paper to read performance differences as paradigm effects rather than implementation noise.","core_discovery":"On the paper's own terms, the discovery is that the apparent chaos of cross-view matching methods resolves into a structured design space plus a reproducible performance ordering. The proposed taxonomy organizes dozens of methods along six orthogonal dimensions and, within matching paradigms, by motivation: accuracy, efficiency, generalization, and distribution. Under the survey's consistent evaluation—SuperPoint for all sparse methods, fixed resizing for semi-dense methods, each dense method at its native settings, RANSAC for pose and homography, and the standard HLoc pipeline for localization—three results stand out. First, on relative pose estimation (MegaDepth, ScanNet) and homography estimation (HPatches), accuracy rises with matching density: dense methods such as RoMa and RoMa v2 lead, cascaded semi-dense methods such as IMAmatch approach them, and sparse methods lag most on weakly textured indoor scenes. Second, sparse methods' ceiling is set by the detector: even strong attention-based matchers collapse on ScanNet because they inherit limited keypoints. Third, in visual localization (Aachen Day-Night, InLoc), semi-dense methods are the best overall, with dense methods losing ground at night and sparse methods falling below 80% indoors.","pith_inferences":["Because dense methods were trained separately on each dataset while sparse and most semi-dense methods used a single MegaDepth-trained model, part of the dense advantage on ScanNet may reflect extra in-domain training; re-running the benchmark with all methods trained on identical data would isolate algorithmic merit from training conditions.","The strong result of cascaded semi-dense methods like IMAmatch suggests the sparse/dense dichotomy is dissolving, and hybrid detector-free pipelines with adaptive density will keep closing the gap to dense methods at lower cost.","If the detector bottleneck is real, replacing SuperPoint with a stronger detector or a detector-free sparse matcher could lift sparse methods on indoor benchmarks more than any further attention redesign.","The nighttime localization results hint that dense methods overfit to day-lit appearance, so testing dense matchers with illumination augmentation or multi-hypothesis matching could recover their localization performance."],"forward_implications":["Keypoint detector quality is the binding constraint on sparse matchers: on weakly textured indoor scenes their pose-estimation AUC drops sharply no matter how strong the attention module is.","Accuracy scales with matching density, so applications that need pose quality in low-texture environments should prefer dense or cascaded semi-dense models, accepting their larger parameter counts.","Semi-dense matching is currently the best default for visual localization, combining indoor robustness with outdoor day/night reliability, while dense methods are more vulnerable to low-light noise.","Foundation models are best used as coarse semantic or geometric guides inside specialized matching networks, not as wholesale replacements, because their features are too coarse for precise localization.","The survey's listed open problems—uncertainty-aware matching, explicit geometric reasoning at inference, non-rigid scenes, and quadratic-cost global matching—define the next generation of correspondence models."],"supporting_citations":[{"why":"Defines the attention-based sparse matching paradigm that anchors the sparse category in the benchmark.","marker":"[4]"},{"why":"Introduces the coarse-to-fine semi-dense Transformer matcher that anchors the semi-dense category and its evaluation protocol.","marker":"[5]"},{"why":"Anchors the dense category and supplies the widely used coarse-semantic and fine-geometric matching paradigm.","marker":"[10]"},{"why":"Provides the MegaDepth dataset and the MegaDepth-1500 protocol used for outdoor relative pose evaluation.","marker":"[186]"},{"why":"Provides the ScanNet indoor dataset used to test weak-texture robustness; several underlined results come from ScanNet-trained models.","marker":"[187]"},{"why":"Supplies the HPatches benchmark with viewpoint and illumination splits used for homography estimation.","marker":"[190]"},{"why":"Provides the Aachen Day-Night outdoor visual localization benchmark with day and night queries.","marker":"[191]"},{"why":"Provides the InLoc indoor visual localization benchmark used for the indoor localization comparison.","marker":"[192]"},{"why":"Supplies the standard HLoc hierarchical localization pipeline used in all visual localization experiments.","marker":"[193]"},{"why":"NetVLAD retrieval selects candidate reference views inside the localization pipeline before matching.","marker":"[194]"}],"fun_headline_variants":["Dense matchers win accuracy; semi-dense win localization","Sparse matchers capped by detector quality in cross-view test","Unified taxonomy and benchmark for cross-view feature matching","Survey: vision foundation models shift cross-view matching design","Consistent protocols reveal honest cross-view matching rankings"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison is fair enough to rank paradigms; that holds only if results obtained under different training-data conditions are directly comparable, since dense methods are trained separately on each dataset, some semi-dense ScanNet numbers come from ScanNet-trained models (marked by underlining), and parameter counts are missing for several methods without public code.","fun_headline_variants_meta":{"raw":{"variants":["Dense matchers win accuracy; semi-dense win localization","Sparse matchers capped by detector quality in cross-view test","Unified taxonomy and benchmark for cross-view feature matching","Survey: vision foundation models shift cross-view matching design","Consistent protocols reveal honest cross-view matching rankings"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000999,"raw_usage":{"total_tokens":4266,"prompt_tokens":1021,"completion_tokens":3245,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":637,"completion_tokens_details":{"reasoning_tokens":3166}},"tokens_in":637,"tokens_out":3245,"duration_ms":19991,"temperature":1.0,"reasoning_tokens":3166,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:25:37.206811+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain every sparse and semi-dense method on ScanNet's own training split, the same data dense methods see, and re-run relative pose estimation. If the dense advantage at AUC@20° shrinks below the gap reported in Table I, or if a sparse method overtakes a dense one, the claimed paradigm hierarchy reflects training conditions rather than matching density.","supporting_citations":[{"cited_title":"MegaDepth: Learning single-view depth pre- diction from internet photos,","cited_arxiv_id":null,"evidence_quote":"Provides the MegaDepth dataset and the MegaDepth-1500 protocol used for outdoor relative pose evaluation."},{"cited_title":"ScanNet: Richly-annotated 3D reconstructions of indoor scenes,","cited_arxiv_id":null,"evidence_quote":"Provides the ScanNet indoor dataset used to test weak-texture robustness; several underlined results come from ScanNet-trained models."},{"cited_title":"HPatches: A benchmark and evaluation of handcrafted and learned local descrip- tors,","cited_arxiv_id":null,"evidence_quote":"Supplies the HPatches benchmark with viewpoint and illumination splits used for homography estimation."},{"cited_title":"Bench- marking 6DoF outdoor visual localization in changing conditions,","cited_arxiv_id":null,"evidence_quote":"Provides the Aachen Day-Night outdoor visual localization benchmark with day and night queries."},{"cited_title":"InLoc: Indoor visual localization with dense matching and view synthesis,","cited_arxiv_id":null,"evidence_quote":"Provides the InLoc indoor visual localization benchmark used for the indoor localization comparison."},{"cited_title":"From coarse to fine: Robust hierarchical localization at large scale,","cited_arxiv_id":null,"evidence_quote":"Supplies the standard HLoc hierarchical localization pipeline used in all visual localization experiments."},{"cited_title":"NetVLAD: CNN architecture for weakly supervised place recognition,","cited_arxiv_id":null,"evidence_quote":"NetVLAD retrieval selects candidate reference views inside the localization pipeline before matching."}],"review_version":1}