{"id":"492f3508-1537-400a-a3b6-6e9f66e03e7f","arxiv_id":"2607.03972","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"DS-SAC deterministically estimates robust geometric models via residual-density search and recursive partitioning, outperforming RANSAC variants in AUC and speed on large vision datasets.","lead":"DS-SAC is a deterministic robust estimator that finds geometric models by searching dense residual regions with forward/backward percentile refinement and recursive signed-residual partitioning, instead of random minimal samples. On large multi-dataset benchmarks it reports higher pose AUC and lower runtime than common RANSAC variants for homography and epipolar geometry.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The density-plus-signed-split heuristic may fail when the full-set LS start is outlier-dominated, so reported AUC gains may not generalize beyond the tested SuperPoint+LightGlue regime.","rationale":"The Reader correctly isolates the density-plus-signed-split premise as the weakest assumption and assigns CONDITIONAL because code is unreleased and error bars are absent. My stress-test sharpens the same premise: the full-set LS initialization (explicit in Alg. 2) is the concrete point of failure when outliers dominate, and the paper supplies no stress test of that regime. The proposed injection experiment would settle whether the reported AUC gains survive lower inlier ratios; until then the verdict remains CONDITIONAL with the same high confidence. No stronger internal inconsistency or formal error was found.","tokens_in":14937,"tokens_out":548,"duration_ms":5382,"concrete_test":"Re-run the exact Table 3 (fundamental-matrix) protocol on the same 39 592 pairs after injecting controlled random mismatches that lower the SuperPoint+LightGlue inlier ratio in 10 % steps down to 20 %. If DS-SAC’s AUC@10° advantage over GC-RANSAC disappears or reverses below ~40 % inliers while runtime remains low, the density-start assumption is load-bearing and the headline claim must be qualified.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim (higher AUC + faster runtime vs RANSAC/MAGSAC/LO-/GC-RANSAC on Tables 2–4) rests on the §3 observation that “inliers form a dense region in the residual space of a good model,” so that percentile shrinking from a full-set least-squares start (Alg. 2, θ_init = arg min_θ Σ d²) plus signed-residual bipartition (Alg. 4) will locate high-consensus models. When the initial support is heavily contaminated, that LS start is biased; successive percentile selection then tracks a dense outlier cluster rather than the true inliers, and signed splits (r = x′⊤Fx or the analogous homography residual) merely recurse on the same biased residual field. The paper never measures or bounds the inlier ratio at which the heuristic still recovers the global consensus; all 39 592 pairs use SuperPoint+LightGlue matches whose residual density already favors the method. Ablation (Table 5) and sensitivity (Figs. 1–2) only vary internal parameters, not the quality of the initial residual landscape. Consequently the empirical superiority may be an artifact of the particular matcher rather than a general property of density search.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes DS-SAC, a deterministic robust estimator for geometric models (homography, fundamental matrix, essential matrix). Instead of RANSAC-style random minimal sampling, it initializes a model by least-squares on the current point set, then performs forward search (shrinking percentile support and re-fitting) and backward search (expanding from the local best), with recursive bipartition of the support by the sign of a model residual (Algorithms 1–4). Model selection uses inlier count with MSAC as tie-breaker, plus a final multi-threshold post-tuning stage. The authors give a binary-tree complexity argument of O(N²) worst-case and O(N log N) balanced, and report higher pose AUC, competitive/lower median pose error, and lower wall-clock time than OpenCV RANSAC, MAGSAC, LO-RANSAC, and GC-RANSAC on 39,592 SuperPoint+LightGlue pairs from six public datasets (Tables 2–4), with sensitivity and ablation studies (Figs. 1–2, Table 5).","tokens_in":15240,"tokens_out":1339,"duration_ms":20873,"significance":"If the empirical gains hold under broader conditions, DS-SAC is a practically useful deterministic alternative to stochastic consensus methods for standard multi-view geometry tasks: it avoids iteration-count tuning, has polynomial (not exponential-in-outlier-ratio) cost, and is faster than strong LO/GC-RANSAC baselines while improving AUC. Strengths that should be credited include the large multi-dataset, multi-model evaluation protocol aligned with IMC-style pose AUC, the explicit complexity analysis, the ablation of backward search and post-tuning, and the commitment to release code. The work sits between LO-style residual refinement and expensive global consensus maximizers; that positioning is clear and of interest to the geometric vision community.","major_comments":[{"comment":"§3 (motivating observation) and Algorithm 2: the method starts from θ_init = arg min_θ Σ_{x_i∈S} d²(x_i,θ) on the full current partition, then shrinks percentiles of residual-ordered points. The central claim that density search reliably finds high-consensus models therefore depends on the residual landscape of this (possibly heavily contaminated) least-squares start still having a dense inlier basin that percentile shrinking can track. The manuscript never measures the inlier-ratio regime in which this holds: all 39,592 pairs use SuperPoint+LightGlue, and neither synthetic outlier sweeps nor weaker matchers appear. Without such stress tests (or a bound), the superiority in Tables 2–4 may not generalize beyond the tested residual-density regime, which is load-bearing for presenting DS-SAC as a general alternative to RANSAC.","section":"§3, Algorithm 2, Tables 2–4"},{"comment":"§5.2 Experimental Setup: stochastic baselines are fixed at 1000 iterations while DS-SAC runs an adaptive number of full least-squares refits (~465–490 reported). Wall-clock times favor DS-SAC, which is useful, but the accuracy comparison is not an equal-compute or equal-hypothesis budget study, and MAGSAC’s reported times (0.278–0.402 s) are outliers relative to the others. A short equal-time or equal-iteration-cost protocol (or reporting of inlier-ratio stratified results) is needed to support the claim of consistently higher AUC at lower cost.","section":"§5.2, Tables 2–4"},{"comment":"§4 Computational Complexity: the O(N²) worst-case derivation sets Δp = 1/N and builds an extremely unbalanced tree of height N−1. Operating parameters are Δp = 0.03 and p_min = 0.2, under which the number of refits is far smaller and roughly linear in the number of partitions. The analysis should be restated for the actual parameter regime used in experiments (and ideally matched to the observed ~480 iterations), so that the “polynomial and efficient” claim is not left resting on an unrealistic step size.","section":"§4"}],"minor_comments":[{"comment":"Notation table and Algorithms 2–3: P vs X vs S is used inconsistently for the full set vs current partition; Algorithm 2 writes ∀x_i ∈ P while the input is S. Unify symbols.","section":"Table 1, Algorithms 2–3"},{"comment":"Algorithm 4 exception handling for partition boundary outside the current space is described in text but omitted from the pseudocode; a short note or branch would aid reproducibility.","section":"§3.3, Algorithm 4"},{"comment":"Figs. 1–2 and Tables 2–4 report point estimates only; even simple per-dataset standard deviations or bootstrap intervals on AUC would strengthen the multi-dataset average claim.","section":"§5.3–5.6"},{"comment":"Related work could more explicitly contrast residual-percentile search with PROSAC’s quality-ordered sampling and with LO-RANSAC’s inlier re-estimation, to clarify novelty beyond “deterministic LO-style refinement.”","section":"§2"},{"comment":"Minor typos/grammar: “bases on the sign” (Alg. 1), “Nist´ er”, “re-scan the space in the forward direction without the inlier optimization step” could be tightened for clarity.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The empirical package is solid for a methods paper and the density-search idea is worth publishing if the generalization concern is addressed. I would not reject on novelty alone—the combination of percentile density search + signed residual recursion + poly-time claim is distinct enough from LO/GC-RANSAC—but without outlier-ratio or matcher stress tests the central “deterministic alternative” claim remains conditional on modern dense matchers. Scope fits a solid CV journal; major revision is the right bar."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a clean algorithmic paper that delivers a usable deterministic alternative to RANSAC for homography, fundamental, and essential matrix estimation. The core idea is new enough to matter: start from a full-set least-squares fit, then do percentile-based forward shrinking and backward expansion of the residual support, followed by recursive bipartition on the sign of the residual (Sampson or the analogous homography residual). That combination is not in LO-RANSAC, GC-RANSAC, MAGSAC, or the older branch-and-bound consensus literature. Complexity is polynomial (O(N^{2}) worst-case, O(N log N) balanced), which is a real selling point.\n\nThe experiments are the strongest part. They run the same SuperPoint+LightGlue matches on ~40k pairs from six public datasets and report consistent AUC gains at 5/10/20° plus lower or competitive median pose error and clearly lower runtime than OpenCV RANSAC, MAGSAC, LO-RANSAC, and GC-RANSAC under a fixed 1000-iteration budget. Ablations (Table 5) and sensitivity plots (Figs. 1–2) show the design choices are not brittle inside the tested range. MSAC tie-breaking and the final post-tuning step are sensible. The math is elementary and correctly stated; the citation pattern is fair and covers the right prior art.\n\nThe soft spot is exactly the one the stress-test flags, and it is real but not fatal. The method assumes that inliers form a dense residual region reachable by successive percentile shrinking from a full-set LS start. When that start is heavily outlier-dominated the search can lock onto a dense outlier cluster and the signed splits just recurse on the same biased field. The paper never measures the inlier-ratio threshold at which this happens, and every number is generated with SuperPoint+LightGlue matches that already produce relatively clean residual landscapes. So the reported superiority may not transfer to weaker matchers or pathological scenes. No error bars or statistical tests either, and code is only promised. Those are ordinary conference-paper limitations, not load-bearing contradictions.\n\nThis is for people who ship multi-view geometry pipelines and want a deterministic, fast estimator they can actually drop in. It is not a foundational advance, but it is solid engineering with honest empirical support. I would send it to peer review; a referee can push on the density assumption and ask for a couple of harder matchers. Worth reading and, if the code appears, worth trying.","headline":"Practical deterministic RANSAC drop-in that wins AUC and speed on the three standard multi-view tasks; density heuristic works on modern matchers but its failure boundary is unmapped.","tokens_in":15808,"tokens_out":618,"would_cite":true,"duration_ms":11331,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A deterministic residual-density search finds better geometric models than random sampling, faster.","keywords":["RANSAC","robust estimation","sample consensus","homography","fundamental matrix","essential matrix","deterministic consensus","residual density search"],"falsifier":"On the same SuperPoint+LightGlue correspondences and fixed-budget protocol, if DS-SAC's AUC at 5/10/20 degrees and median pose error on ScanNet1500, PhotoTourism, LaMAR, 7Scenes, ETH3D, and KITTI fail to beat or match LO-RANSAC and GC-RANSAC for homography, fundamental, and essential matrices, the central empirical claim fails.","tokens_in":15824,"feed_emoji":"📐","tokens_out":1043,"duration_ms":11770,"temperature":0.7,"pith_summary":"Robust model fitting from noisy point matches still usually means RANSAC: guess tiny random subsets until one looks good. This paper argues that you can drop the randomness. Start from a model fit to the points you have, then repeatedly shrink and expand the set of nearest residuals (forward and backward search) so the estimate walks into denser regions of residual space; when local search stalls, split the points by the sign of their residuals and recurse on each half. The resulting method, DS-SAC, has polynomial cost in the number of points and, on tens of thousands of real image pairs for homography and epipolar geometry, reports higher pose-recall AUC, competitive or lower median pose error, and lower runtime than standard RANSAC variants under a fixed iteration budget. A sympathetic reader cares because many vision pipelines are bottlenecked by fragile, slow consensus steps; a deterministic, density-driven alternative that is both more accurate and cheaper would change how those pipelines are built.","feed_headline":"Skip random sampling: dense residual search beats RANSAC","feed_subtitle":"DS-SAC raises pose-recall AUC and cuts runtime on homography and epipolar geometry across 39k pairs.","key_machinery":"Density Search Sample Consensus (DS-SAC): forward/backward percentile residual search plus recursive signed-residual partitioning. It replaces stochastic minimal sampling with deterministic local density refinement and residual-space bipartition so that inlier-dense support sets are located without exhaustive model-space search.","core_discovery":"DS-SAC establishes that high-consensus geometric models can be found deterministically by searching dense residual regions: initialize from all points in a partition, refine by percentile-based forward then backward residual selection plus inlier re-estimation, and globally explore by recursive bipartition on signed residuals until a minimum partition size, finally selecting by inlier count with MSAC tie-break. On large multi-dataset benchmarks this yields higher AUC at 5/10/20 degrees, competitive or better median pose error, and faster runtimes than RANSAC, MAGSAC, LO-RANSAC, and GC-RANSAC for homography, fundamental, and essential matrix estimation.","pith_inferences":["The same residual-density and signed-split pattern may transfer to other algebraic models (e.g., trifocal tensor or plane-and-parallax) that admit a signed residual and a least-squares solver.","If percentile step and minimum partition size can be scheduled adaptively from residual histograms, iteration count could fall further without losing the dense-region guarantee.","Hybrid use—DS-SAC as a deterministic seed followed by a short LO or graph-cut polish—could combine its global residual exploration with spatial coherence that pure density search does not encode.","Failure modes when multiple dense clusters of similar residual magnitude exist (structured outliers) would stress-test whether bipartition alone separates the true consensus set."],"forward_implications":["Homography, fundamental, and essential matrix estimation can use a fixed deterministic search instead of random minimal samples under high outlier ratios.","Runtime for consensus on large match sets can drop while pose-recall AUC rises, because search cost is polynomial rather than exploding with outlier rate.","Local optimization no longer needs a lucky RANSAC hypothesis as a trigger; residual-density refinement is the whole procedure.","MSAC can serve only as a secondary tie-break once inlier count is maximized, keeping the objective aligned with classical consensus.","Polynomial complexity (O(N^2) worst case, O(N log N) when balanced) makes the method a practical alternative where exhaustive deterministic consensus is too expensive."],"fun_headline_variants":["DS-SAC skips random sampling for dense residual consensus","Deterministic density search tops RANSAC on geometry tasks","Residual density search finds models faster than RANSAC","Recursive residual partitions beat stochastic consensus","DS-SAC: polynomial dense search for high-consensus models"],"cache_read_input_tokens":128,"weakest_assumption_plain":"Inliers of a good model form a dense enough region in residual space that shrinking from a full-set least-squares start, then splitting by residual sign and repeating, will find high-consensus models without exhaustive search.","fun_headline_variants_meta":{"raw":{"variants":["DS-SAC skips random sampling for dense residual consensus","Deterministic density search tops RANSAC on geometry tasks","Residual density search finds models faster than RANSAC","Recursive residual partitions beat stochastic consensus","DS-SAC: polynomial dense search for high-consensus models"]},"model":"grok-4.5","effort":"low","cost_usd":0.002924,"raw_usage":{"total_tokens":1078,"prompt_tokens":790,"num_sources_used":0,"completion_tokens":62,"cost_in_usd_ticks":29240000,"prompt_tokens_details":{"text_tokens":790,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":226,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":790,"tokens_out":62,"duration_ms":2492,"temperature":1.0,"reasoning_tokens":226,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T22:37:11.057595+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On the same SuperPoint+LightGlue correspondences and fixed-budget protocol, if DS-SAC's AUC at 5/10/20 degrees and median pose error on ScanNet1500, PhotoTourism, LaMAR, 7Scenes, ETH3D, and KITTI fail to beat or match LO-RANSAC and GC-RANSAC for homography, fundamental, and essential matrices, the central empirical claim fails.","supporting_citations":[],"review_version":1}