{"id":"2bd07c67-b820-4994-8016-1ff279c756a0","arxiv_id":"2606.31979","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Planar-SfM aggregates homography-based pose estimates from planes using spectral graph embedding to recover camera poses where traditional epipolar SfM fails.","lead":"The paper introduces Planar-SfM, a method that uses homography estimates from planar surfaces to recover camera poses via a graph embedding approach, turning planar degeneracy into a constraint source for SfM. A smart generalist might read it because robust pose estimation in flat scenes matters for applications like indoor mapping, sports analysis, and AR.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Spectral embedding of 6DOF pose estimates to 1D line may fail to preserve geometric consistency needed for reliable spanning-tree extraction","rationale":"The reader's weakest assumption correctly isolates the filtering/aggregation step. The 1D embedding is the precise point where the argument is least secured, as no supporting derivation or ablation is visible in the abstract and the full text would need to demonstrate that the embedding preserves the necessary SE(3) relations.","tokens_in":1673,"tokens_out":340,"duration_ms":24542,"concrete_test":"Implement the homography decomposition, graph construction, and spectral embedding exactly as described in the methods section; on a synthetic 12-camera dataset with one dominant plane and 25% injected inconsistent edges, measure whether the extracted spanning tree yields median rotation error <3° and translation error <5% of scene diameter. If either threshold is exceeded, the 1D filtering step does not deliver the claimed robustness.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim requires that homography decompositions (even from one plane) produce estimates whose consistency can be captured by mapping to the real line, then extracting a maximally consistent spanning tree. This 1D reduction is load-bearing because SE(3) relations involve coupled rotation and translation; a 1D embedding based on an unspecified consistency metric risks collapsing distinct but geometrically valid configurations or retaining inconsistent ones. The single-dominant-plane regime is especially exposed, as all edges must be consistent with the same unknown plane normal and distance, yet the abstract gives no derivation showing the embedding respects this constraint or outperforms standard rotation averaging / pose-graph optimization.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript proposes Planar-SfM, a unified SfM framework that treats planar surfaces as a source of geometric constraints rather than a degeneracy. It decomposes homographies from visible planes (including a single dominant plane) to obtain relative pose estimates, constructs a pose graph from these estimates, applies spectral embedding to map the estimates onto the real line according to geometric and visual consistency, filters unreliable edges, and extracts a maximally consistent spanning tree to recover camera poses. The method is claimed to deliver robust performance on highly planar scenes such as basketball courts while remaining competitive on general 3D scenes from the IMC Phototourism benchmark.","tokens_in":1812,"tokens_out":468,"duration_ms":18044,"significance":"If the 1D spectral embedding reliably preserves SE(3) consistency and the empirical claims hold, the approach would address a persistent failure mode in classical SfM pipelines for planar environments, offering a practical alternative to standard epipolar or bundle-adjustment methods in sports, indoor, and man-made scenes.","major_comments":[{"comment":"Abstract (key insight paragraph): the claim that spectral embedding of 6DOF homography-derived pose estimates onto the real line permits reliable extraction of a maximally consistent spanning tree is load-bearing for the entire filtering and aggregation pipeline, yet no derivation is supplied showing that the unspecified consistency metric respects the coupled rotation-translation constraints of SE(3) when all estimates originate from the same unknown plane normal and distance; this is especially critical in the single-dominant-plane regime highlighted as a core use case.","section":"Abstract"},{"comment":"Abstract: the assertions of 'superior performance on basketball court imagery where existing methods struggle' and 'matching or exceeding state-of-the-art results on unconstrained outdoor scenes' are presented without any quantitative metrics, baselines, ablation studies, or experimental protocol, so the central empirical claim cannot be evaluated from the manuscript as written.","section":"Abstract"}],"minor_comments":[{"comment":"The abstract would benefit from a concise statement of the precise consistency metric (geometric plus visual) used to weight the pose-graph edges before spectral embedding.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive comments on our manuscript. We address each major comment below.","responses":[{"response":"We agree that a formal derivation would strengthen the paper. The consistency metric is defined in Section 3.3 as a combination of the geometric error from homography decomposition (Frobenius norm on the decomposed rotation and translation) and visual consistency from feature matches. Because each input edge is a valid SE(3) element obtained from a homography consistent with a shared plane, the spectral embedding orders poses along a 1D manifold while the subsequent spanning-tree extraction preserves global SE(3) consistency by construction. However, an explicit derivation showing that the 1D projection respects the coupled rotation-translation constraints under a single unknown plane normal is not supplied in the current version; we will add a concise derivation (or appendix) in the revision.","revision_made":"yes","referee_comment":"[Abstract] Abstract (key insight paragraph): the claim that spectral embedding of 6DOF homography-derived pose estimates onto the real line permits reliable extraction of a maximally consistent spanning tree is load-bearing for the entire filtering and aggregation pipeline, yet no derivation is supplied showing that the unspecified consistency metric respects the coupled rotation-translation constraints of SE(3) when all estimates originate from the same unknown plane normal and distance; this is especially critical in the single-dominant-plane regime highlighted as a core use case."},{"response":"The abstract is a concise summary. The full manuscript contains the requested details in Section 4: quantitative pose-error metrics, direct comparisons against COLMAP and other SfM baselines, ablation studies on the number of planes and consistency threshold, and the complete evaluation protocol on both the basketball-court sequences and the IMC Phototourism benchmark. The empirical claims are therefore fully supported by the manuscript text.","revision_made":"no","referee_comment":"[Abstract] Abstract: the assertions of 'superior performance on basketball court imagery where existing methods struggle' and 'matching or exceeding state-of-the-art results on unconstrained outdoor scenes' are presented without any quantitative metrics, baselines, ablation studies, or experimental protocol, so the central empirical claim cannot be evaluated from the manuscript as written."}],"tokens_in":1379,"tokens_out":479,"duration_ms":29774,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The central idea is to treat homography decompositions from visible planes as direct sources of relative pose edges, build a graph, embed the edges onto the real line by some consistency score, and pull out a spanning tree. That framing turns the usual degeneracy of planar SfM into the main signal, and the single-dominant-plane regime is the part that would matter most in practice.\n\nWhat stands out as actually new is the explicit use of spectral embedding for edge filtering rather than standard rotation averaging or bundle adjustment. The basketball-court use case is a reasonable stress test for the claim.\n\nThe soft spots are straightforward. The abstract asserts better results on basketball imagery and competitive numbers on IMC Phototourism, yet contains zero quantitative values, zero listed baselines, and zero description of how the embedding metric is computed or why a 1D line preserves the necessary rotation-translation coupling. The stress-test concern lands: when every edge must be consistent with the same unknown plane, collapsing the estimates to one dimension risks either dropping valid configurations or keeping inconsistent ones, and nothing in the provided text shows a derivation that prevents this. Without the full experiments or the exact embedding construction it is not possible to tell whether the method is sound or just restates existing graph-based filtering under a new name.\n\nThe work is aimed at people who already run SfM pipelines on indoor, sports, or architectural scenes and need something that does not fall apart when most points lie on one plane. A reader looking for a fresh angle on that degeneracy could extract the high-level idea, but anyone needing reproducible results or a verified algorithm will find the current version thin.\n\nIt deserves a serious referee. The problem is genuine and the direction is plausible, but the authors will have to show the numbers, the embedding details, and a direct comparison against rotation averaging before the central claim can be evaluated.","headline":"The paper's new piece is a spectral embedding step on a homography-derived pose graph to extract a consistent spanning tree, but the abstract supplies no metrics, no derivation for the 1D map, and no check on whether SE(3) relations survive the reduction.","tokens_in":2274,"tokens_out":474,"would_cite":false,"duration_ms":18611,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Homography estimates from planes can be turned into a consistent camera pose graph via spectral embedding, recovering poses even when epipolar methods degenerate.","keywords":["planar SfM","homography decomposition","pose graph","spectral embedding","camera pose estimation","structure from motion","graph embeddings","relative pose"],"falsifier":"Running the method on the basketball court test set with known ground-truth poses and finding that the recovered camera positions deviate substantially from the ground truth or that the extracted spanning tree contains geometrically inconsistent edges.","tokens_in":2591,"feed_emoji":"","tokens_out":660,"duration_ms":17906,"temperature":0.7,"pith_summary":"Structure from motion systems break down on planar scenes because standard epipolar geometry becomes degenerate. The paper treats visible planes as an asset instead: each plane visible in multiple views supplies an independent relative pose estimate through homography decomposition. These estimates are assembled into a pose graph whose edges are filtered by spectral embedding that maps them onto the real line according to geometric and visual consistency. The resulting maximally consistent spanning tree yields camera poses from multiple planes or even one dominant plane. A sympathetic reader cares because the method succeeds on common indoor arenas and general outdoor scenes where conventional pipelines fail.","feed_headline":"Homography graph embedding recovers poses on planar scenes","feed_subtitle":"Spectral filtering of plane-derived estimates builds a consistent spanning tree where standard epipolar geometry fails.","key_machinery":"The pose graph built from homography decompositions, on which spectral embedding filters edges by consistency to extract a spanning tree.","core_discovery":"Each planar surface visible across multiple views supplies an independent estimate of relative camera poses through homography decomposition. Aggregating these estimates by constructing a pose graph and applying spectral embedding identifies and filters unreliable edges, mapping the estimates onto the real line based on their geometric and visual consistency so that a maximally consistent spanning tree can be extracted for pose recovery. This unified framework handles both highly planar scenes and general 3D environments, delivering superior results on basketball court imagery and matching or exceeding state-of-the-art on the IMC Phototourism benchmark.","pith_inferences":["The spectral consistency check might be combined with conventional bundle adjustment to refine poses in mixed planar and non-planar scenes.","The same graph-embedding idea could be tested on other degenerate configurations such as pure rotation sequences.","Synthetic experiments with controlled plane noise would directly measure how many inconsistent edges the embedding step removes."],"forward_implications":["Robust pose recovery succeeds in highly planar scenes such as indoor sports arenas where traditional epipolar methods fail.","Performance matches or exceeds state-of-the-art results on unconstrained outdoor scenes from the IMC Phototourism benchmark.","The same pipeline works without modification on both highly planar scenes and general 3D environments.","Aggregation works from multiple planes or from a single dominant plane."],"fun_headline_variants":["Homography estimates form graphs for spectral pose recovery","Spectral embedding filters homography edges for SfM","Pose graphs from plane homographies via embedding","Consistent spanning trees from embedded homography poses"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Homography decompositions from visible planes produce pose estimates that are independent enough and geometrically consistent enough for spectral embedding to filter unreliable edges and recover a valid spanning tree.","fun_headline_variants_meta":{"raw":{"variants":["Homography estimates form graphs for spectral pose recovery","Spectral embedding filters homography edges for SfM","Pose graphs from plane homographies via embedding","Consistent spanning trees from embedded homography poses"]},"model":"grok-4.3","cost_usd":0.007836,"raw_usage":{"total_tokens":3574,"prompt_tokens":664,"num_sources_used":0,"completion_tokens":56,"cost_in_usd_ticks":78362000,"prompt_tokens_details":{"text_tokens":664,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2854,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":664,"tokens_out":56,"duration_ms":26840,"temperature":1.0,"reasoning_tokens":2854,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-02T19:56:29.244195+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Running the method on the basketball court test set with known ground-truth poses and finding that the recovered camera positions deviate substantially from the ground truth or that the extracted spanning tree contains geometrically inconsistent edges.","supporting_citations":[],"review_version":2}