{"id":"64560ff7-fbc4-4f14-a729-15c802c7a2c8","arxiv_id":"2608.09112","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"ROEVO organizes image edge pixels into ordered chains, then uses only those organized edges for RGB-D camera tracking and local bundle adjustment, beating several point-, edge-, and learning-based methods in sparse-texture benchmarks.","lead":"This paper presents ROEVO, a visual odometry system that turns scattered image edge pixels into ordered chains called organized edges, then uses those chains for camera tracking and local mapping. If it works as reported, it offers a practical way to get accurate RGB-D camera poses in sparse-texture indoor and outdoor scenes where point-feature methods struggle.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sparse-texture robustness is argued from an ICL-NUIM protocol that deletes exactly the bare-wall frames where edge-only tracking would be stressed; this weakens the central claim more than any threshold detail.","rationale":"The reader's weakest assumption (edge sequentialization and 5° normal association) is plausible and worth testing; if those thresholds are fragile, associations corrupt and the pipeline degrades. I would put the ICL-NUIM truncation ahead of it as the load-bearing issue because it is explicitly disclosed, directly targets the central claim, and is easy to test: the paper removes the frames that would stress the thresholds. The manuscript has real independent support in the form of broad experiments across ICL-NUIM, TUM, ETH-3D, chair, and plant-fence data, an ablation table, and a promised open-source release, so I do not see the work as unsound. My concern does not move the verdict: the paper remains a solid engineering contribution whose central robustness claim is conditional on evaluation on the hardest sequences. Independent re-running on untruncated ICL-NUIM (and release of the promised code) would settle it.","tokens_in":28453,"tokens_out":8784,"duration_ms":100683,"concrete_test":"Run ROEVO-TR and ROEVO-TR+BA on the untruncated ICL-NUIM lr_kt0–lr_kt3 sequences, plus at least ORB-SLAM3-LM and Manhattan-SLAM under the same protocol. Report ATE, RPE, and the fraction of frames in which tracking is lost or the per-frame pose error exceeds a pre-registered threshold. If a substantial fraction of the previously excluded bare-wall frames cause failure or large drift, the claim should be weakened from 'robust in sparse-texture scenes' to 'robust where extractable edges exist'.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is robustness in sparse-texture scenes. Table I (ATE 0.633 cm on ICL-NUIM) is the main supporting evidence. But §VII-B.1 truncates every sequence segment where the viewpoint 'only captures bare walls' and excludes 'frames where pose estimation is not feasible'; the reported average is a frame-count-weighted mean over surviving segments. That removes the exact failure mode the claim is about: with no Canny edges, organized-edge extraction, oriented-BFS sequentialization, 5° normal association, and the co-visibility graph all have nothing to operate on. The comparison to baselines on the truncated inputs may be fair, but it cannot establish robustness in sparse-texture scenes if the hardest sparse-texture frames are discarded. The association thresholds flagged by the reader are a real fragility, but they would manifest as drift or loss on these hard frames; the protocol prevents that evidence from being seen. A sensitivity study of the 20°/5° thresholds is still worth doing, but the omission of unlocalizable frames is the more direct threat to the headline claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"ROEVO is an RGB-D visual odometry system that uses only edge features. Edge pixels are clustered into sequentially ordered 'organized edges' via oriented BFS, then tracked with a coarse photometric step and a fine edge-wise association step that combines geometric point-to-tangent and photometric residuals. A local mapping thread maintains a co-visibility graph via a disjoint set, fuses associated edges into 3D map edges, and performs a decoupled bundle adjustment. The system is evaluated on ICL-NUIM, TUM RGB-D, ETH-3D, and two custom datasets, with ablations for coarse tracking, repeatability, and runtime, and it reports state-of-the-art ATE on many sequences.","tokens_in":28713,"tokens_out":7305,"duration_ms":75683,"significance":"The organized-edge representation is a useful and well-motivated contribution: it gives edge features an ordered structure that supports edge-level association, co-visibility, and local map fusion, and the empirical study is extensive, with public code. If the reported results hold, the paper demonstrates that an edge-only VO can be competitive with point-based, direct, and learning-based methods, particularly in structure-rich but texture-sparse scenes. The main limitation is that the sparse-texture claim is weakened by the ICL-NUIM evaluation protocol, and the core optimization equations contain inconsistencies that must be corrected.","major_comments":[{"comment":"The ICL-NUIM protocol truncates segments that 'only capture bare walls' and excludes 'frames where pose estimation is not feasible,' then reports a frame-count-weighted average. This removes exactly the frames in which an edge-only system has no edge features to associate, so Table I cannot support the paper's headline claim of robustness in sparse-texture environments. Please report results on the full sequences, including failure counts or lost-track segments, and compare baselines on the same full inputs, or restrict the claim to segments where edges are actually observable.","section":"Section VII-B.1, Table I"},{"comment":"Equation (8) has a dimension mismatch: e_g is in R^2, p' is a 2D reprojected point, and the projection Jacobian in Eq. (25) is 2x6, so the left factor must be I_{2x2} - uu^T rather than I_{3x3} - ... . In Eq. (12), the Hessian terms for the photometric part omit the alpha^2 factor that follows from J_a = [J_g; alpha J_p] in Eq. (9), while Eq. (13) contains alpha^2; H and g are therefore inconsistent. These equations are central to the fine-tracking and BA implementation and need to be corrected and re-derived.","section":"Section V-C, Eqs. (8)-(9) and (12)-(13)"},{"comment":"The transformation of joint BA into 'argmin_T (argmin_E ...)' in Eq. (18) is not an equality for general nonlinear least squares; the authors acknowledge that the subsequent use of edge fusion is approximate ('approximately preserve the minimum fitting geometric residuals'). Since the paper presents decoupled BA as a contribution, this approximation should be stated clearly, justified with a convergence or accuracy comparison against joint BA on a subset of sequences, and its error quantified. As written, the derivation implies an exactness that the method does not provide.","section":"Section VI-C, Eqs. (17)-(19)"},{"comment":"The extraction and association depend on two fixed thresholds: 20 degrees for gradient-direction clustering in oriented BFS and 5 degrees for edge-normal association in fine tracking. No sensitivity analysis is provided, and these thresholds directly determine whether edges are sequentialized and whether correspondences survive viewpoint changes. Please add a sensitivity study (e.g., 5-40 degrees and 1-20 degrees) on at least TUM and ICL sequences, and report how often tracking is lost as the thresholds vary.","section":"Section IV-B and Section V-C"}],"minor_comments":[{"comment":"The first paragraph contains corrupted text: 'pose esfine time:18.6219 timates' should be cleaned up.","section":"Section I"},{"comment":"The header uses 'RMES' instead of 'RMSE', and the average columns repeat 'RMES (t)'; please correct these typos.","section":"Table II"},{"comment":"The caption repeats the ICL-NUIM sentence '(c) and (d) are the estimated trajectories ... lr kt2 and lr kt1' in the ETH-3D results; replace it with the actual ETH-3D sequence names.","section":"Fig. 13 caption"},{"comment":"The row for ManhattanSLAM cites reference [54], but the correct reference is [34].","section":"Table IX"},{"comment":"The ablation table reports a single row of RPE values without naming the sequences or indicating how the values are aggregated; add the sequence list and standard deviations.","section":"Section VII-C, Table XII"},{"comment":"The loop-closure error on the Plant-Fence sequence is reported on a single trajectory without ground truth; state that this is a qualitative indicator and report repeated runs if available.","section":"Table XIII"},{"comment":"The pseudo-code uses informal conventions such as 'set union' and 'associate list[j] != false'; define these operations or replace them with standard mathematical notation.","section":"Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the ICL-NUIM truncation rather than the threshold details: the paper's central robustness claim is not fully supported by the surviving-segment evaluation. The Jacobian and Hessian inconsistencies in Section V-C are fixable but must be corrected, and the authors should verify that the public code matches the corrected equations. I would support acceptance after these points are addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the good news: the organized-edge idea is genuinely new. Turning Canny pixels into sequentialized chains via oriented BFS with gradient clustering, then associating whole edges by normal consistency and building a co-visibility graph with a disjoint set, is not in the cited CannyVO, RE-SLAM, or quadtree work. The paper builds a full VO around that representation, with a coarse-to-fine tracker, edge fusion, and a decoupled BA, and tests it across ICL-NUIM, TUM, ETH-3D, and self-collected data, including ablations and repeatability experiments. Code is promised. That is a solid engineering contribution, and the empirical breadth is real.\n\nThe soft spots are addressable but need attention. The fine-tracking Jacobian in Eq. 8 is dimensionally inconsistent: e_g in Eq. 7 is a 2D vector in the image plane, yet J_g is written with a 3x3 matrix. The appendix gives the correct 2x6 projection Jacobian, so the printed formula does not type-check. Easy fix, but it needs fixing.\n\nThe BA decomposition in Eq. 18 is an acknowledged approximation: they substitute edge-fusion centroids for true fitting. That is fine as an engineering choice, but it means the \"joint optimization\" is really alternating fitting and pose-only registration, not BA in the usual sense. The claim of shape preservation should be read as approximate.\n\nThe more serious concern is the ICL-NUIM protocol. The paper truncates segments where the viewpoint only captures bare walls and reports a frame-weighted mean over surviving segments. This is a fair comparison for all baselines, but it removes exactly the frames where an edge-only tracker would be stressed. So the headline \"robust in sparse-texture scenes\" is not actually demonstrated on the hardest sparse-texture images; the ETH-3D and chair results help, but the ICL evidence is weaker than the abstract suggests.\n\nAlso minor: the 20-degree clustering and 5-degree association thresholds are fixed, with no sensitivity study. That is worth adding but not disqualifying.\n\nWho is this for? Anyone working on edge-based VO/SLAM, direct methods, or RGB-D odometry will get value from the representation and the experimental comparisons. The paper deserves a serious referee if the authors fix the Jacobian, clarify the ICL truncation as a limitation, and ideally add threshold sensitivity. I would engage with it, and I would bring it to the reading group.","headline":"A genuinely new organized-edge representation and a complete VO built on it, with a strong empirical effort, but the ICL-NUIM truncation protocol undercuts the headline robustness claim and the fine-tracking Jacobian has a dimension slip.","tokens_in":29253,"tokens_out":2487,"would_cite":true,"duration_ms":26378,"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":"Organized edges—ordered chains of edge pixels—let a visual odometry system track and map using only edges, outperforming point-feature and learning-based baselines in sparse-texture scenes.","keywords":["organized edges","visual odometry","RGB-D cameras","edge-based tracking","bundle adjustment","co-visibility graph","sparse texture","edge features"],"falsifier":"Take a sequence with sharp corners or significant motion blur, extract organized edges with the paper's fixed thresholds, and count how many true scene edges fragment into multiple chains or fail to associate across viewpoints; if the fragmentation rate is high enough that the co-visibility graph inherits wrong correspondences, the reported robustness in sparse-texture scenes would collapse.","tokens_in":1445,"feed_emoji":"📐","tokens_out":2746,"duration_ms":72208,"temperature":0.7,"pith_summary":"This paper claims that image edges, which most visual odometry systems treat as loose pixels, can be turned into ordered chains it calls organized edges, and that this representation alone is enough to build a full RGB-D visual odometry system that tracks and maps accurately. The authors argue that previous edge-based methods fail because they scatter edge pixels, losing the sequential order that carries structural information. By clustering edges by gradient direction and ordering each cluster into a single path, the system can associate whole edges across frames, build a co-visibility graph, and run bundle adjustment over fused map edges. The payoff, if true, is a VO system that stays accurate in sparse-texture and structure-heavy indoor scenes where point features and learning-based pipelines struggle.","feed_headline":"Edges, ordered into chains, beat point features in visual odometry","feed_subtitle":"New edge-only VO system matches or beats state-of-the-art on sparse-texture datasets.","key_machinery":"The load-bearing object is the organized edge: a Canny edge mask is preprocessed to remove two ambiguous local pixel patterns, then traversed by an oriented breadth-first search that clusters adjacent pixels whose gradient directions differ by at most 20 degrees, recording the expansion route. Backtracking from end nodes and retaining the longest combined path turns each cluster into a single ordered sequence of pixels. This ordering is what makes edge-level association possible: normals and tangents are computed from neighboring points in the chain, strict 5-degree normal-direction agreement validates candidate correspondences, and a disjoint-set structure chains associations across keyframes into a co-visibility graph. The same ordering supports the fusion of multiple views of one spatial edge into a fitted 3D map edge, and the decoupled bundle adjustment that alternates fitting with per-pose registration.","core_discovery":"The central discovery is that representing edges as organized, sequentialized clusters rather than raw edge masks makes edge-only visual odometry viable and competitive. Each organized edge is an ordered point sequence whose textural information comes from gradient orientation and whose structural information comes from curvature along the chain. With this representation, the paper builds edge-wise residuals for 3D-2D registration, a disjoint-set co-visibility graph for multi-frame association, a shape-preserving fitting and stitching step for map edges, and a bundle adjustment that decouples into edge fitting followed by per-pose registration. On ICL-NUIM, TUM RGB-D, ETH-3D, and custom indoor and outdoor sequences, the full system achieves the best ATE rank in 13 of 24 sequences and the top-three rank in 18 of 24, while completing sparse-texture sequences where several baselines fail.","pith_inferences":["Beyond the paper's experiments, the fixed 20-degree clustering threshold implies an upper bound on detectable curvature: sharp corners will fragment into several organized edges, so the stitching step in edge fusion likely carries much of the burden in man-made scenes with many right angles.","A natural testable extension is to make the 20-degree and 5-degree thresholds adaptive to local edge curvature or gradient noise; the paper's own note that motion blur remains a weakness suggests such adaptation could widen the operating range.","The organized-edge representation could plausibly serve as a constraint source inside a full SLAM system with loop closure, since it already produces a co-visibility graph and fused map elements that look like landmarks.","Because the association is edge-level rather than pixel-level, the representation may transfer to monocular or stereo settings, where depth comes from triangulation instead of a depth sensor; the paper lists this as future work, not a demonstrated result."],"forward_implications":["An RGB-D VO system that uses only edges can run at roughly 35 Hz on a laptop CPU and remain accurate in sparse-texture indoor rooms.","Sliding-window bundle adjustment over organized edges reduces rotation drift and improves local mapping quality, as shown by the relative pose error improvements from tracking-only to tracking-plus-BA configurations.","Because organized edges encode gradient orientation, coarse tracking with a rotation-invariant patch-based photometric error beats raw Canny edges and keypoint-selection baselines in relative pose accuracy.","The method completes sequences where distance-field edge methods and geometry-based pipelines fail, indicating that the representation transfers to cluttered outdoor scenes with repeated structures and dynamic shadows.","Edge fusion filters pseudo-edges caused by dynamic disturbances, so the semi-dense reconstruction in scenes with human shadows is cleaner than the raw edge input would allow."],"supporting_citations":[{"why":"Supplies the Canny edge detector whose binary mask is the raw input for organized-edge extraction.","marker":"[28]"},{"why":"Provides the sparse direct method and keypoint-selection baseline that motivates the coarse photometric tracking on edges.","marker":"[1]"},{"why":"Serves as a baseline edge-based VO using distance fields, whose limitations motivate the organized-edge representation.","marker":"[20]"},{"why":"Serves as a baseline edge-based SLAM using distance fields and is a comparison target for robustness and failure cases.","marker":"[21]"},{"why":"Supplies a line-and-plane based baseline that the paper must outperform in structure-rich indoor scenes.","marker":"[34]"},{"why":"Provides the point-feature-based SLAM baseline used in tracking-only and local-mapping modes for comparison and ablation.","marker":"[45]"},{"why":"Provides the disjoint-set data structure used to build the multi-frame co-visibility graph.","marker":"[46]"},{"why":"Inspires the decoupling of bundle adjustment into edge fitting and per-pose registration, which the paper adapts to organized edges.","marker":"[48]"},{"why":"Supplies the TUM RGB-D benchmark used to evaluate tracking and mapping accuracy against real-world baselines.","marker":"[47]"},{"why":"Supplies the synthetic ICL-NUIM benchmark used to test the method in sparse-texture indoor scenes.","marker":"[50]"}],"fun_headline_variants":["Chained edges make edge-only visual odometry viable","Organized edges turn sparse textures into navigable paths","Edge chains give visual odometry an edge over points","Sequentialized edges boost VO accuracy on low-texture scenes"],"cache_read_input_tokens":31360,"weakest_assumption_plain":"The whole pipeline assumes that true scene edges, after Canny extraction and suppression of two ambiguous local patterns, can be unambiguously sequentialized by oriented BFS with a fixed 20-degree gradient-direction clustering threshold, and that such chains repeat across viewpoints enough for the 5-degree normal-association threshold.","fun_headline_variants_meta":{"raw":{"variants":["Chained edges make edge-only visual odometry viable","Organized edges turn sparse textures into navigable paths","Edge chains give visual odometry an edge over points","Sequentialized edges boost VO accuracy on low-texture scenes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000181,"raw_usage":{"total_tokens":1325,"prompt_tokens":981,"completion_tokens":344,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":279}},"tokens_in":597,"tokens_out":344,"duration_ms":4166,"temperature":1.0,"reasoning_tokens":279,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:16:32.226709+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a sequence with sharp corners or significant motion blur, extract organized edges with the paper's fixed thresholds, and count how many true scene edges fragment into multiple chains or fail to associate across viewpoints; if the fragmentation rate is high enough that the co-visibility graph inherits wrong correspondences, the reported robustness in sparse-texture scenes would collapse.","supporting_citations":[{"cited_title":"A linear-time algorithm for a special case of disjoint set union,","cited_arxiv_id":null,"evidence_quote":"Provides the disjoint-set data structure used to build the multi-frame co-visibility graph."},{"cited_title":"Canny-vo: Visual odometry with rgb-d cameras based on geometric 3-d–2-d edge alignment,","cited_arxiv_id":null,"evidence_quote":"Serves as a baseline edge-based VO using distance fields, whose limitations motivate the organized-edge representation."},{"cited_title":"Reslam: A real-time robust edge- based slam system,","cited_arxiv_id":null,"evidence_quote":"Serves as a baseline edge-based SLAM using distance fields and is a comparison target for robustness and failure cases."},{"cited_title":"Manhattanslam: Robust planar tracking and mapping leveraging mixture of manhattan frames,","cited_arxiv_id":null,"evidence_quote":"Supplies a line-and-plane based baseline that the paper must outperform in structure-rich indoor scenes."},{"cited_title":"A benchmark for the evaluation of rgb-d slam systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the TUM RGB-D benchmark used to evaluate tracking and mapping accuracy against real-world baselines."},{"cited_title":"A benchmark for rgb-d visual odometry, 3d reconstruction and slam,","cited_arxiv_id":null,"evidence_quote":"Supplies the synthetic ICL-NUIM benchmark used to test the method in sparse-texture indoor scenes."}],"review_version":1}