{"id":"c0f983ac-654b-41b0-a82d-5c92bba55e44","arxiv_id":"2506.23897","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A dual-branch optical flow network using a 90-degree rotated 'orthogonal' view reduces polar distortion errors and sets new state-of-the-art results on MPFDataset and FlowScape.","lead":"PriOr-Flow is a new dual-branch neural network for panoramic optical flow that also runs the image pair through a 90-degree rotated spherical view to compensate for projection distortion near the poles. It reports state-of-the-art accuracy on two public panoramic flow benchmarks, with the largest gains in the high-distortion polar regions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never defines how a vector flow field transforms between primitive and orthogonal views; Eq. 12 and Eq. 18 reuse the image-space map T for flow fields, and ODDC's Eq. 13 confidence inherits this ambiguity.","rationale":"The reader's weakest assumption targets Eq. 13's use of primitive-view group-wise correlation as a proxy for ortho-branch reliability after re-projection. That is a valid concern, but it presupposes that the re-projection itself is correct. The more fundamental, load-bearing gap is that the paper never specifies how a vector displacement field is transformed under the spherical rotation: Equations 12 and 18 reuse the same symbol T defined for points/images, while flow vectors require a different composition involving the inverse projection. If the implementation resamples displacement arrays, then both the ortho supervision and the ODDC confidence are evaluated in the wrong coordinate system, and the claimed polar-region gains would not support the stated mechanism. This concern is concrete and falsifiable: the authors release code, so a unit test with an analytically known spherical motion can settle it in either direction. I do not treat this as evidence of misbehavior; it is an unverified definitional step. The paper's strengths remain: public code, architecture-agnostic integration, and consistent large gains on two datasets, plus qualitative real-world results. Those strengths justify keeping the reader's CONDITIONAL verdict: the empirical claim is plausible, but the central coordinate transformation must be verified before the mechanism can be accepted as stated. Hence UNCHANGED.","tokens_in":13543,"tokens_out":13069,"duration_ms":150954,"concrete_test":"Write a unit test around the released implementation of T^p_o for flow fields: for a synthetic panoramic pair with a known 3D rotation (e.g., a pure rotation about the z-axis), compute the ground-truth primitive flow, convert it to the ortho view with the code's transformation, then convert back and compare to the analytic primitive flow P(Y)-P(X) at every pixel. If the discrepancy exceeds about 1 pixel away from seams, the re-projection in Eq. 12/18 is not vector-correct; then rerun the Table 1 ablations with a corrected transformation and check whether the polar EPE gain persists. If the unit test passes, replace Eq. 13's Go2p with an ortho-domain group-wise correlation computed on f^o_1, f^o_2 with F^o and mapped to primitive, and compare the resulting polar EPE in Table 1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism depends on converting the ortho branch's flow field into primitive coordinates: Eq. 12 uses T^p_o(F^o) for ODDC, and Eq. 18 supervises the ortho branch with T^o_p(F_gt). However, Section 3.1 only defines T for points/images via spherical rotation and bilinear interpolation (Eq. 3), not for displacement vectors. A flow vector is not a point on the sphere-image grid: expressing an ortho flow at a primitive pixel requires composing the ortho-pixel lookup with the inverse spherical projection, i.e., F^o2p(x) = P(O^{-1}(O(X)+F^o(O(X)))) - x, not resampling the displacement array. If T^p_o is implemented as resampling, then F^o2p in Eq. 12 is not the primitive-coordinate representation of the ortho flow, Eq. 13's Go2p measures primitive feature similarity under the wrong displacement, and the ortho branch is trained against inconsistent targets in Eq. 18. The large polar EPE reduction attributed to ODDC in Table 1 (7.56 to 5.57) would then be an artifact of a coordinate error rather than evidence for the low-distortion prior. This assumption is unstated and load-bearing, but it is checkable because the code is public.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PriOr-Flow, a dual-branch framework for panoramic optical flow estimation. The primitive branch operates on the original equirectangular view, while a second branch operates on an orthogonal view obtained by a 90-degree spherical rotation, which has low distortion near the poles. Two components are introduced: the Dual-Cost Collaborative Lookup (DCCL), which retrieves correlation cues from both cost volumes, and the Ortho-Driven Distortion Compensation (ODDC) module, which fuses motion features from both branches using confidence maps. The method is integrated with RAFT, GMA, and SKFlow. Experiments on MPFDataset and FlowScape report state-of-the-art EPE and SEPE, with particularly large gains in polar regions, and the code is publicly available.","tokens_in":13716,"tokens_out":4732,"duration_ms":56065,"significance":"If the reported results are correct, the paper makes a substantial empirical contribution: it is the first panoramic optical flow method that explicitly targets polar-region distortion through an orthogonal-view branch, and it demonstrates large, consistent improvements over prior SOTA on two public benchmarks. The proposed modules are architecture-agnostic, and the ablations in Tables 3 and 5 suggest that the method is broadly applicable and even effective with fewer iterations. The public code release is a strength, and the paper is honest about the equatorial trade-off in Table 7. However, the central mechanism depends on a coordinate transformation for flow fields that is not defined in the paper; this must be resolved before the significance of the empirical gains can be assessed.","major_comments":[{"comment":"","section":"Section 3.1, 3.4, 3.5 (Eqs. 3, 12, 18)"},{"comment":"","section":"Section 3.2, Eq. (7)"},{"comment":"","section":"Section 4.3, Tables 1, 2, 4, 5"}],"minor_comments":[{"comment":"","section":"Section 3.1, Eq. (3)"},{"comment":"","section":"Section 3.4, Eq. (13)"},{"comment":"","section":"Tables 1-7"}],"recommendation":"major_revision","confidential_remarks":"The flow-field transformation issue is the key technical risk. If the code implements the correct endpoint-based transformation and the paper's notation is merely imprecise, the paper is a strong candidate after revision. If the implementation resamples displacement arrays, the reported polar gains may be artifacts of the coordinate mismatch; the public code will quickly show which case holds. The test-set selection concern in Section 4.3 is secondary but should be addressed for the benchmark claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Core idea is clean: rotate the ERP frame 90 degrees about the x-axis, run a second flow branch on the rotated view, and fuse the two cost volumes. The orthogonal view has low distortion where the primitive has high distortion, so the fusion supplies the polar regions with better correlation cues. That is genuinely new—prior multi-view work stitched or simply concatenated flows—and the authors back it with consistent gains on three base architectures (RAFT, GMA, SKFlow). They are also honest about the equatorial trade-off in Table 7. Public code is a real plus.\n\nThe main technical worry is the flow vector transformation. The paper defines T for points and images (Eq. 3), then uses the same map on flow fields (Eqs. 12 and 18). A flow vector is not a point; resampling the displacement field is not the same as transforming the endpoint and subtracting the transformed start. If the implementation does the latter correctly, fine—but the paper does not say. If it only resamples, then the ortho branch is trained against a wrong target and the ODDC confidence is computed with a misaligned displacement. That could explain part of the polar gain, and the stress-test note is right to flag it. It is checkable in the posted code, and the authors need to clarify.\n\nThe other soft spots are more mundane: no error bars or repeated-trial statistics; the rotation axis/angle and iteration count are selected on the EFT test split; and the strongest prior baseline PanoFlow is absent from the MPF dataset comparison. All are fixable with modest experiments, and none by itself undermines the central idea.\n\nFor a reader: this paper is worth discussing for the multi-view fusion mechanism, and certainly worth citing if the transformation question gets resolved. My guess is a competent referee can settle it quickly. Recommend sending it to review with a request to spell out exactly how F^o2p is computed and to add validation-based selection and a PanoFlow comparison on MPF.","headline":"A genuinely new orthogonal-view fusion idea for panoramic flow, with big reported gains, but the flow-field transformation is under-specified and must be clarified before the numbers can be trusted.","tokens_in":14370,"tokens_out":2897,"would_cite":true,"duration_ms":31402,"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":"PriOr-Flow claims optical flow in equirectangular panoramas improves by processing an orthogonally rotated view of the same sphere, cutting reported endpoint error by about 30 percent on two benchmarks.","keywords":["panoramic optical flow","equirectangular projection","orthogonal view","dual-branch network","cost volume lookup","polar region distortion","iterative refinement"],"falsifier":"Take the best model and replace ODDC's learned confidence with either (a) the fixed distortion map of the primitive ERP view or (b) confidence computed in each branch's own coordinates; if polar-region EPE on FlowScape does not get worse under (a) or does not improve under (b), the mechanism attributed to ODDC is not carrying the gain. A second decisive experiment: train the full model with the orthogonal branch's loss term removed. If the 29.6% EPE reduction on FlowScape persists without supervision on the orthogonal branch, the low-distortion prior itself is not the cause, and the improvement instead comes from the shared lookup or the extra parameters.","tokens_in":13233,"feed_emoji":"🌐","tokens_out":11588,"duration_ms":98653,"temperature":0.7,"pith_summary":"The paper sets out to show that the severe polar-region errors of equirectangular (ERP) panoramic optical flow can be removed by processing a second, orthogonally rotated copy of the same sphere alongside the original. Because a 90-degree rotation puts the second view's minimum distortion exactly where the original has its maximum, the network can harvest low-distortion motion evidence for the poles instead of guessing through the distortion. This matters for any application built on wide-field motion — virtual reality, autonomous driving, video interpolation and inpainting — because ERP remains the standard panoramic format. The best variant, PriOr-RAFT, is reported to reduce endpoint error by 30.0% on MPFDataset and 29.6% on FlowScape relative to the previous state of the art, with the polar regions contributing most of the gain.","feed_headline":"Panoramic flow error drops 30% with a rotated twin view","feed_subtitle":"A second panorama, rotated 90 degrees, supplies low-distortion cues exactly where equirectangular poles fail.","key_machinery":"The load-bearing object is the orthogonal view: the same spherical scene re-projected into ERP after a 90-degree rotation about the x-axis, whose distortion pattern is complementary to the primitive view's, with minimal distortion exactly at the poles. Two mechanisms built on it carry the argument. DCCL is a lookup operator that indexes both cost volumes at spherical locations derived from a single flow estimate, so correlation cues from the low-distortion orthogonal volume suppress the distortion noise in the primitive volume's polar regions. ODDC is a fusion module that estimates both branches' confidence as the group-wise correlation between frame-1 features and warped frame-2 features, re-projects the orthogonal flow into primitive coordinates, and uses the confidences to gate how the orthogonal motion features compensate the primitive branch inside a ConvGRU. A spherical-area-weighted L1 loss supervises both branches so the non-uniform ERP sampling does not bias training toward the equator.","core_discovery":"The central claim is that the orthogonal view's distortion pattern is complementary to the primitive ERP view's — a 90-degree rotation about the x-axis sends the primitive view's most distorted content (the poles) to the equator, so the rotated view has minimal distortion exactly where the original is worst — and that this complementarity can be converted into a working training signal. Two mechanisms do the conversion. The Dual-Cost Collaborative Lookup (DCCL) operator takes the current flow estimate, maps it onto the sphere, and retrieves correlation scores from both the primitive and the orthogonal cost volumes at the same spherical locations, so noisy polar-region cues in one volume are countered by clean cues in the other. The Ortho-Driven Distortion Compensation (ODDC) module estimates a per-pixel confidence for each branch as the group-wise correlation between frame-1 features and frame-2 features warped by that branch's flow, converts the orthogonal flow back into primitive coordinates, and feeds confidences together with both correlation cues into a confidence-guided ConvGRU, so the primitive flow is corrected adaptively exactly where its own evidence is weak. Trained end-to-end with a spherical-area-weighted loss on both branches, the framework reaches state-of-the-art endpoint error on both public benchmarks, including a 39.7% polar-region improvement over PanoFlow on FlowScape.","pith_inferences":["Extension: the complementary-distortion trick is not specific to flow; stereo disparity, monocular depth, and video super-resolution on ERP input all suffer pole errors and could plausibly use the same rotate-and-fuse design.","Part of the gain may be a two-view data effect: the dual-branch design supervises two views of the same motion with shared lookups, effectively doubling the training signal. An ablation that trains the primitive branch alone on the same frames, with the orthogonal branch present at inference but unsupervised, would separate the low-distortion-prior contribution from the extra-supervision contribut","ODDC's confidences are computed in primitive pixel coordinates only; computing confidence in each branch's own coordinates and fusing on the sphere might close the small equatorial-region gap the authors report against PanoFlow.","Real-world support is currently qualitative on OmniPhotos and ODVista; a quantitative test with pseudo-ground-truth from a depth sensor or a multi-camera rig would show whether the synthetic-dataset gains survive real stitching and lens artifacts."],"forward_implications":["Because DCCL and ODDC are grafted onto the iterative refinement loop, the framework transfers: RAFT, GMA, and SKFlow backbones all improve on the EFT scene of MPFDataset, with EPE gains from 15.8% to 23.7% at full iterations.","The accuracy gain concentrates in the poles: on FlowScape the polar-region EPE is 39.7% better than PanoFlow, while the equatorial EPE is comparable (slightly worse), so the orthogonal prior is what buys the headline reduction.","The orthogonal prior accelerates convergence: PriOr-RAFT with only 4 iterations already beats the 12-iteration RAFT baseline by 13.4% EPE, and 3 iterations beat RAFT at the same count.","The 90-degree x-axis rotation is the load-bearing viewpoint choice: ablations show a y-axis rotation splits the polar region into a discontinuity, and 45-degree x-rotation leaves parts of the poles uncompensated, both performing worse."],"supporting_citations":[{"why":"Supplies the all-pairs cost volume and iterative ConvGRU refinement that DCCL and ODDC extend; it is also the primary baseline and backbone.","marker":"[41]"},{"why":"Contributes the MPFDataset benchmark and the spherical-area-weighted loss used to supervise both branches.","marker":"[28]"},{"why":"Provides the FlowScape dataset, the leading competitor on it, and the equatorial-versus-polar baseline against which the 39.7% polar gain is measured.","marker":"[35]"},{"why":"The previous state of the art on MPFDataset that PriOr-RAFT beats by 30.0% EPE, representing the siamese-representation approach.","marker":"[6]"},{"why":"Defines the group-wise correlation operation used in Eq. 13 to compute the confidence maps in ODDC.","marker":"[18]"},{"why":"The tangent-plane baseline method whose projection strategy is compared and contrasted with the orthogonal-view approach.","marker":"[48]"},{"why":"One of the iterative perspective backbones into which PriOr-Flow is integrated to demonstrate universality.","marker":"[22]"},{"why":"Another backbone used in the universality experiments.","marker":"[39]"}],"fun_headline_variants":["Rotated twin view cuts polar flow error by 39.7%","Orthogonal view rescues polar regions in panoramic flow","Dual-view lookup fixes distortion for panoramic flow","Polar flow improved 39.7% by adding a rotated view"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"ODDC decides how much to trust each branch by measuring group-wise correlation in primitive-view pixel coordinates only, so the whole polar-region gain rests on the assumption that a flow that is accurate on the sphere still looks accurate after being re-expressed in the primitive ERP grid — if that re-projection breaks the correspondence, the fusion weights would be systematically wrong.","fun_headline_variants_meta":{"raw":{"variants":["Rotated twin view cuts polar flow error by 39.7%","Orthogonal view rescues polar regions in panoramic flow","Dual-view lookup fixes distortion for panoramic flow","Polar flow improved 39.7% by adding a rotated view"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000912,"raw_usage":{"total_tokens":3956,"prompt_tokens":1024,"completion_tokens":2932,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":640,"completion_tokens_details":{"reasoning_tokens":2862}},"tokens_in":640,"tokens_out":2932,"duration_ms":21238,"temperature":1.0,"reasoning_tokens":2862,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:29:02.390769+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the best model and replace ODDC's learned confidence with either (a) the fixed distortion map of the primitive ERP view or (b) confidence computed in each branch's own coordinates; if polar-region EPE on FlowScape does not get worse under (a) or does not improve under (b), the mechanism attributed to ODDC is not carrying the gain. A second decisive experiment: train the full model with the orthogonal branch's loss term removed. If the 29.6% EPE reduction on FlowScape persists without supervision on the orthogonal branch, the low-distortion prior itself is not the cause, and the improvement instead comes from the shared lookup or the extra parameters.","supporting_citations":[{"cited_title":"Deep 360° optical flow estimation based on multi- projection fusion","cited_arxiv_id":null,"evidence_quote":"Contributes the MPFDataset benchmark and the spherical-area-weighted loss used to supervise both branches."},{"cited_title":"Panoflow: Learning 360° optical flow for surrounding tem- poral understanding","cited_arxiv_id":null,"evidence_quote":"Provides the FlowScape dataset, the leading competitor on it, and the equatorial-versus-polar baseline against which the 39.7% polar gain is measured."},{"cited_title":"Learning omnidirectional flow in 360 video via siamese representation","cited_arxiv_id":null,"evidence_quote":"The previous state of the art on MPFDataset that PriOr-RAFT beats by 30.0% EPE, representing the siamese-representation approach."},{"cited_title":"Group-wise correlation stereo network","cited_arxiv_id":null,"evidence_quote":"Defines the group-wise correlation operation used in Eq. 13 to compute the confidence maps in ODDC."},{"cited_title":"360 optical flow us- ing tangent images","cited_arxiv_id":null,"evidence_quote":"The tangent-plane baseline method whose projection strategy is compared and contrasted with the orthogonal-view approach."},{"cited_title":"Learning to estimate hidden motions with global motion aggregation","cited_arxiv_id":null,"evidence_quote":"One of the iterative perspective backbones into which PriOr-Flow is integrated to demonstrate universality."},{"cited_title":"Skflow: Learning optical flow with super kernels","cited_arxiv_id":null,"evidence_quote":"Another backbone used in the universality experiments."}],"review_version":1}