{"id":"807d1ffd-4674-4d5a-a3d2-21232499376d","arxiv_id":"1908.05547","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Local patch descriptors learned on log-polar sampled patches match keypoints across up to 4x scale mismatch and beat Cartesian-patch baselines on multiple benchmarks.","lead":"This paper shows that sampling image patches in a log-polar pattern instead of the usual rectangular grid lets learned image descriptors tolerate scale mismatches between matched keypoints. The result could make 3D reconstruction and visual localization more robust to scale estimation errors.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Scale-invariance gains are quantified only on COLMAP-derived correspondences; no independent scale-controlled evaluation supports the 2-4x claim.","rationale":"The reader's weakest assumption pointed to COLMAP-derived correspondence noise; I agree and sharpen it: the specific quantitative claim about 2-4x scale robustness is only evaluated on that pipeline. The independent HPatches/AMOS/PhotoTourism results support a general advantage but do not isolate scale ratio, so they cannot rescue the wide-range number. The occlusion part of the abstract is also asserted rather than tested; Table 2's lambda sweep is consistent with occlusion/background robustness but does not control occlusions. A secondary literature issue: Section 2 states 'there is no learning-based method that explicitly addresses scale invariance' while citing [19] (Scale-Aware Triplet Networks), which by title and abstract is such a method; the novelty claim should be revised, though this does not affect the experimental evidence. Overall the paper's empirical contribution is strong enough for conditional acceptance, which is the reader's verdict; I recommend no movement from CONDITIONAL.","tokens_in":16188,"tokens_out":13083,"duration_ms":92919,"concrete_test":"On a dataset with known homographies (e.g., HPatches or a synthetic textured scene), detect SIFT keypoints in each image and randomly perturb their scale estimates by fixed multiplicative factors 1.5x, 2x, 3x, and 4x while keeping locations and orientations fixed. Extract LogPol lambda=96, Cart lambda=12/96, and baseline descriptors, and compute FPR95 per scale-factor bin using homography-verified correspondences rather than COLMAP depth maps. If the LogPol advantage at 3-4x falls below the paper's 'negligible drop' standard or becomes comparable to Cart, the Section 4.1.2 claim is pipeline-dependent. Report results with error bars across multiple image pairs.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The quantitative heart of the central claim—'negligible drop in performance under scale changes up to 2-3x, and remain useful even at 3-4x' (Section 4.1.2, Fig. 4)—is measured on the authors' own test set, whose ground-truth correspondences are produced by the same COLMAP pose/depth pipeline (Section 4.1.1) that generated the training data, with a 1.5-pixel projection threshold and cyclic consistency check. If COLMAP depth errors are spatially correlated with appearance, the positive pairs in the test set could be biased toward cases that log-polar handles for reasons other than genuine scale invariance, e.g., dominance of low-frequency central content. The HPatches and AMOS experiments are independent and do show gains, but they are not stratified by scale mismatch, so they do not substantiate the specific 2-4x range. The scale-ratio histogram in Fig. 4 is also computed from the same depth/pose estimates, so errors in those estimates directly contaminate the x-axis of the claimed result. An independent scale-controlled benchmark is required to confirm that the wide-range claim is not an artifact of the evaluation pipeline.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes replacing the standard Cartesian sampling of local patch support regions with direct log-polar sampling via a Polar Transformer Network (PTN), followed by a HardNet-style CNN that produces 128-dimensional descriptors. The authors build a new training set from YFCC photo-tourism images processed with COLMAP, deliberately collecting SIFT keypoint pairs whose scales are not matched, and train with a hardest-in-batch triplet loss. They report FPR95 on their own test set, rank-1 retrieval on HPatches and AMOS, and pose-estimation mAP on the CVPR 2019 PhotoTourism challenge. The central claims are that log-polar descriptors tolerate scale mismatches up to 2-4x with negligible degradation, that they can exploit much larger support regions (up to lambda=96) without the performance collapse seen for Cartesian patches, and that this yields state-of-the-art results on three datasets.","tokens_in":16462,"tokens_out":6832,"duration_ms":66893,"significance":"If the claims hold, the paper makes a useful and potentially influential contribution: it decouples descriptor matching from the accuracy of detector scale estimation and shows that a simple change in sampling geometry, combined with standard deep learning, can enlarge effective support regions. Strengths include the public release of code, models, and training data; the controlled lambda ablation that keeps the architecture and training pipeline identical between Cartesian and log-polar variants; and validation on three external benchmarks (HPatches, AMOS, PhotoTourism), which goes beyond a single in-house evaluation. The principal weakness is that the quantitative 2-4x scale-invariance claim is established almost entirely on an internally generated test set whose ground-truth correspondences and scale-ratio histogram derive from the same COLMAP depth/pose estimates used to create the training data; an independent, scale-controlled evaluation is needed to make the central claim fully credible.","major_comments":[{"comment":"The central claim that log-polar models show 'a negligible drop in performance under scale changes up to 2-3x, and remain useful even at 3-4x' is evaluated only on the authors' COLMAP-based test set. Ground-truth correspondences in Section 4.1.1 are generated by projecting SIFT keypoints through COLMAP depth maps and estimated poses with a 1.5-pixel threshold and a cyclic consistency check, and the scale-ratio histogram in Fig. 4 is computed from these same depth/pose estimates. If COLMAP depth errors are correlated with appearance or with scale ratio, the positive pairs in the largest-scale bins could be systematically easier for log-polar patches for reasons unrelated to genuine scale invariance. Because this directly supports the abstract's 'much wider range of scales' claim, the authors should add an independent scale-controlled evaluation, for example using synthetic images with known homographies and zoom factors, or stratifying HPatches by the scale ratio induced by the ground-truth homography, and report matching accuracy per scale-ratio bin.","section":"Section 4.1.2 and Fig. 4"},{"comment":"The abstract and Section 4.1.2 state that log-polar support regions can be made much larger 'without suffering from occlusions,' but no experiment directly measures occlusion robustness. Table 2 shows that log-polar FPR95 improves with lambda while Cartesian degrades, yet this is measured on the general COLMAP test set, where a larger lambda also exposes more scale variation and more background content; the improvement could be driven by scale equivariance rather than by occlusion tolerance. An explicit occlusion experiment (for example, masking or occluding controlled fractions of the support region, or reporting performance separately for keypoints with known occlusion masks) is required to substantiate this load-bearing claim.","section":"Abstract; Section 4.1.2 and Table 2"},{"comment":"No error bars, repeated runs, or statistical significance tests are reported for any FPR95 or rank-1 numbers. For example, in Table 3 the viewpoint split shows Ours-LogPol lambda=96 at 0.847 versus lambda=64 at 0.849, and several baseline differences are on the order of 0.001-0.01, so wording such as 'performance increases with lambda, until it saturates' and some comparative statements could depend on variance. Reporting per-sequence standard deviations across training runs or a significance test over sequences would make the empirical conclusions more robust, especially where the reported gaps are small.","section":"Tables 1-4"}],"minor_comments":[{"comment":"The expression e^{log(r_i)} in Eq. (1) should be simplified to r_i; as written it introduces a redundant exponential/log pair and obscures the intended radial coordinate.","section":"Section 3.1, Eq. (1)"},{"comment":"The second line of Eq. (2) writes y_t = y_i + x_s sin(theta_i) sigma_i/W + y_s sin(theta_i) sigma_i/H; the second sine should likely be a cosine to represent a standard rotation in the Cartesian sampler. Please check and correct the formula.","section":"Section 3.1, Eq. (2)"},{"comment":"The text refers to COLMAP outputs as 'ground truth camera poses' and later to 'ground truth correspondences'; since these are estimated quantities, the wording should be softened to 'estimated poses and depth' to avoid implying independent ground truth.","section":"Section 4.1.1"},{"comment":"The abstract's 'state-of-the-art results on three different datasets' overstates Table 5, where the method ranks second on both PhotoTourism tracks, although it is first by average rank. Consider describing the result as 'top-performing on average' or 'second on both tracks' for accuracy.","section":"Abstract and Section 4.4, Table 5"},{"comment":"Several scale/orientation bins are sparsely populated, as the caption acknowledges. Adding sample counts per bin, or suppressing bins with very few matches, would help readers judge which parts of the 2-4x scale-invariance curve are reliable.","section":"Fig. 4 caption"},{"comment":"The text says 'the small gap between HardNet and Ours-Cartesian,' but the average FPR95 values are 0.98 and 0.72, respectively; 'small relative to the other baselines' would be more accurate.","section":"Section 4.1.2, Table 1"}],"recommendation":"major_revision","confidential_remarks":"I see the paper as potentially publishable after an independent, scale-controlled evaluation of the 2-4x scale-invariance claim and after the occlusion claim is tested directly. The authors' openness with code, models, and data makes such additional experiments feasible. The current version's central quantitative claim rests on a single internally generated benchmark, which is the main risk."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this paper is worth a serious look. It changes the input representation for learned patch descriptors—log-polar sampling of raw pixels instead of Cartesian grids—and shows convincing evidence that this buys real scale robustness. The central idea is genuinely new relative to prior work, which used log-polar grids only for pooling hand-crafted features. The authors also demonstrate a clean ablation: as the support region grows, their log-polar models keep improving while Cartesian models collapse. That is a robust and interesting finding.\n\nThe main soft spot is exactly what the stress-test note flags. The headline numbers—'negligible drop up to 2–3x, useful at 3–4x'—come from a test set built with the same COLMAP poses and depth maps used for training, and the scale-ratio histogram is computed from those same estimates. So the precise range is not independently verified. That said, the paper does not rest solely on that dataset. HPatches and AMOS are independent benchmarks, and the log-polar models win there too, even if those benchmarks don't stratify by scale mismatch. So the qualitative conclusion—log-polar representations help with scale variation—survives the concern. The exact 2–4x envelope should be treated with caution, not dismissed.\n\nTwo smaller issues. First, the occlusion robustness claim in the abstract is supported only indirectly, through the λ ablation; there is no direct occlusion test. That is a minor overreach. Second, no error bars or multiple-seed training runs, which is common in this literature but still limits confidence in the size of the gains.\n\nCitation pattern looks honest. They correctly distinguish their sampling approach from log-polar pooling in SIFT, GLOH, Daisy, etc. Code and data are promised. This is a paper that should go to peer review, and I would expect it to be accepted after addressing the scale-evaluation concern and softening the occlusion language. For anyone working on local features, it's worth citing and worth discussing.","headline":"Log-polar sampling of raw pixels is a genuinely new and well-supported idea; the headline scale range is measured on a self-built, COLMAP-dependent test set, so treat the exact 2-4x numbers as provisional.","tokens_in":16919,"tokens_out":2132,"would_cite":true,"duration_ms":21224,"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":"Sampling image patches in log-polar coordinates rather than on a Cartesian grid lets learned local descriptors match keypoints even when the detected scales differ by a factor of 2-3x, with usable performance at 3-4x.","keywords":["local descriptors","log-polar sampling","scale invariance","keypoint matching","deep metric learning","patch retrieval","structure-from-motion"],"falsifier":"Take image pairs with known homographies and deliberately extract keypoints whose scales differ by 2x and 3x; if the log-polar descriptors fail to match them at the same rate as matched-scale keypoints, the central scale-invariance claim fails. Concretely, compute rank-1 accuracy on such controlled mismatched-scale pairs and compare the degradation pattern with the paper's Fig. 4.","tokens_in":16055,"feed_emoji":"🖼️","tokens_out":6206,"duration_ms":58105,"temperature":0.7,"pith_summary":"This paper argues that the usual Cartesian patch, a small square sampled on a regular grid at the detector's estimated scale, is the bottleneck for learned local descriptors. It proposes instead to warp the raw image around each keypoint with a log-polar sampling pattern before feeding it to a convolutional network. Because log-polar coordinates oversample the keypoint's immediate neighborhood and undersample distant regions, changes in scale become approximate horizontal shifts of the patch rather than wholesale content changes. The authors show that descriptors learned on such patches keep matching accuracy nearly constant under 2-3x scale mismatches and remain useful at 3-4x, while also tolerating much larger support regions than Cartesian patches. If true, this removes the need for accurate a priori scale estimation from keypoint detectors in matching pipelines.","feed_headline":"Warping image patches buys a 3-4x scale tolerance","feed_subtitle":"A log-polar sampling pattern lets one descriptor network match keypoints whose detected scales disagree by several factors.","key_machinery":"The load-bearing object is the log-polar sampling pattern, implemented as a polar transformer sampler. Instead of pooling features computed on a Cartesian grid, as earlier log-polar descriptors did, the sampler warps the raw pixel intensities so that the target patch is indexed by log radius and angle. A rotation of the scene becomes a shift along the angular axis, and a scale change becomes a shift along the radial axis, making mismatched-scale patches look similar both to the eye and to the network. This is what lets the same seven-layer convolutional network with a hardest-in-batch triplet loss learn scale-invariant descriptors without any change to the architecture.","core_discovery":"The central claim is that the representation itself, not the network architecture or training loss, is what confers scale invariance. The authors extract a 32x32 log-polar-warped patch around each SIFT keypoint using a polar transformer sampler, while keeping the same convolutional descriptor architecture and hardest-in-batch triplet loss used for Cartesian patches. On a training set built from real photo-tourism images with ground-truth depth maps, they deliberately match keypoints whose detected scales are not in correspondence. Their log-polar models achieve the best false-positive-rate-at-95%-recall on all test sequences, tolerate scale changes up to 2-3x with negligible drop and 3-4x with usable performance, and improve as the support-region size grows up to a radius 8 times larger than the Cartesian optimum, whereas Cartesian models degrade sharply. The same models transfer without fine-tuning to public benchmarks and rank near the top on a pose-estimation challenge.","pith_inferences":["A natural extension, not pursued in the paper, is to make rotation invariance explicit: the log-polar angular axis suggests that full rotation invariance could be learned or even read off directly, building on the orientation jitter already used in training.","The same warp could be used as a scale estimator: because scale changes become shifts, a small correlation layer could recover the scale ratio between matched patches, turning a known failure mode into a measurable quantity.","The boundary problems observed beyond a support radius of 96 point to a concrete improvement: padding or masking the warped patch rather than relying on mirror-padded images could extend the gains to even larger support regions.","More generally, the result invites testing other coordinate warps, such as affine, cylindrical, or spherical sampling, as the input layer for learned descriptors rather than only log-polar."],"forward_implications":["Scale errors from keypoint detectors stop being fatal: matching can proceed even when the two detected scales differ by 2-3x, without re-detecting or re-scaling keypoints.","Support regions can be 8 times larger in radius, or 64 times larger in area, than the Cartesian optimum, so descriptors can use more context without being derailed by occlusions or background motion.","Exposing a network to mismatched-scale Cartesian patches is not enough; the invariance comes from the log-polar warp itself, implying that future invariant descriptors can be designed at the sampling stage.","The models transfer directly to other benchmarks and to pose estimation, so the gains are not confined to the authors' training data.","The same representation should simplify end-to-end pipelines that currently learn scale detection separately, because scale is encoded as a shift rather than as a parameter to estimate."],"supporting_citations":[{"why":"Defines the SIFT keypoints whose scale and orientation estimates are used for patch extraction.","marker":"[23]"},{"why":"Supplies the polar transformer sampling transformation that warps raw patches into log-polar coordinates.","marker":"[11]"},{"why":"Provides the convolutional descriptor architecture and hardest-in-batch triplet loss used for training.","marker":"[28]"},{"why":"Provides the Cartesian spatial transformer sampler used as the comparison baseline.","marker":"[17]"},{"why":"Supplies the structure-from-motion reconstruction and depth maps used to generate ground-truth correspondences at non-matching scales.","marker":"[34]"},{"why":"Defines the standard discriminative patch-training setup that this paper extends to deliberately mismatched scales.","marker":"[7]"},{"why":"Provides one of the public benchmark datasets with ground-truth homographies used for transfer evaluation.","marker":"[3]"}],"fun_headline_variants":["Log-polar patches give descriptors 3-4x scale slack","Scale-agnostic descriptors via log-polar sampling","Log-polar sampling multiplies scale tolerance by 3-4","Log-polar warping: key to scale-robust descriptors","Beyond Cartesian: log-polar warp improves descriptor matching"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The ground-truth matches used for training come from depth maps and camera poses produced by a reconstruction pipeline, so if those reconstructions are noisy or biased toward certain scales, the measured scale-invariance gains may be partly an artifact of the training data.","fun_headline_variants_meta":{"raw":{"variants":["Log-polar patches give descriptors 3-4x scale slack","Scale-agnostic descriptors via log-polar sampling","Log-polar sampling multiplies scale tolerance by 3-4","Log-polar warping: key to scale-robust descriptors","Beyond Cartesian: log-polar warp improves descriptor matching"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000955,"raw_usage":{"total_tokens":4038,"prompt_tokens":878,"completion_tokens":3160,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":494,"completion_tokens_details":{"reasoning_tokens":3080}},"tokens_in":494,"tokens_out":3160,"duration_ms":22015,"temperature":1.0,"reasoning_tokens":3080,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:09:20.379110+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take image pairs with known homographies and deliberately extract keypoints whose scales differ by 2x and 3x; if the log-polar descriptors fail to match them at the same rate as matched-scale keypoints, the central scale-invariance claim fails. Concretely, compute rank-1 accuracy on such controlled mismatched-scale pairs and compare the degradation pattern with the paper's Fig. 4.","supporting_citations":[{"cited_title":"Distinctive Image Features from Scale- Invariant Keypoints","cited_arxiv_id":null,"evidence_quote":"Defines the SIFT keypoints whose scale and orientation estimates are used for patch extraction."},{"cited_title":"Polar Transformer Networks","cited_arxiv_id":null,"evidence_quote":"Supplies the polar transformer sampling transformation that warps raw patches into log-polar coordinates."},{"cited_title":"Working Hard to Know Your Neighbor’s Margins: Local Descriptor Learning Loss","cited_arxiv_id":null,"evidence_quote":"Provides the convolutional descriptor architecture and hardest-in-batch triplet loss used for training."},{"cited_title":"Spatial Transformer Networks","cited_arxiv_id":null,"evidence_quote":"Provides the Cartesian spatial transformer sampler used as the comparison baseline."},{"cited_title":"Sch ¨onberger and Jan-Michael Frahm","cited_arxiv_id":null,"evidence_quote":"Supplies the structure-from-motion reconstruction and depth maps used to generate ground-truth correspondences at non-matching scales."},{"cited_title":"Discrimi- native Learning of Local Image Descriptors","cited_arxiv_id":null,"evidence_quote":"Defines the standard discriminative patch-training setup that this paper extends to deliberately mismatched scales."},{"cited_title":"Hpatches: A Benchmark and Evaluation of Handcrafted and Learned Local Descriptors","cited_arxiv_id":null,"evidence_quote":"Provides one of the public benchmark datasets with ground-truth homographies used for transfer evaluation."}],"review_version":1}