{"id":"35dd62c2-e737-48d3-a44f-e612db000352","arxiv_id":"2602.14021","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A two-view transformer jointly predicting point positions, scene flow, pose weights, and confidence achieves competitive state-of-the-art 4D reconstruction and 3D point tracking without pose-regression heads or bundle adjustment.","lead":"Flow4R predicts per-pixel 3D positions, camera-relative scene flow, pose weights, and confidence from two images in one pass, then derives camera motion, object motion, and point tracks from those quantities with a weighted least-squares solve. The claim is that this compact flow-centric representation makes a single model competitive with — and on several benchmarks better than — separate reconstruction and tracking systems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pseudo-labeled depth/pose likely dominate rigid-flow supervision; the paper never quantifies the GT-vs-pseudo split, so the claimed unified motion supervision may inherit teacher bias.","rationale":"I agree with the reader's CONDITIONAL verdict and with its weakest-assumption identification. The supervision-source ambiguity is the most load-bearing issue because the central claim of a unified, minimal scene-flow formulation rests on the model learning reliable camera-space flow from the proposed rigid-flow construction. If a large share of the 22 datasets contribute only teacher-derived depth and pose, the rigid-flow labels inherit the teacher's errors and possibly a static-scene bias, which would undermine the claimed dynamic-scene generalization. This is an addressable concern: a per-dataset supervision audit and a restricted training run would settle it. Other issues — the SOTA overstatement in the abstract, missing error bars, no code/weights release, and lack of an ablation for the pose-weight map — are relevant but less load-bearing because they do not directly threaten the formulation's validity. The paper's own concluding remark that scene-flow data scarcity constrains flow quality supports the same concern. I therefore recommend no change to the reader's CONDITIONAL verdict: the formulation is plausible and the math checks out, but the empirical support for the central claim is not fully established without clarifying the supervision mix.","tokens_in":16167,"tokens_out":7241,"duration_ms":73885,"concrete_test":"Publish a per-dataset supervision table and retrain using only verified GT depth/pose plus GT motion annotations, then compare WorldTrack APD3D; a large drop would show pseudo-labels are load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on the training supervision in Secs. 1.2 and 3.3: for static scenes, rigid flow is computed from ground-truth depth and relative pose; for dynamic scenes, the same rigid-flow loss LFv (Eq. 14) and the pose-weight loss LW (Eq. 13) require GT depth and camera pose. Section 4 lists 22 training datasets but never states which provide true GT depth/pose, which provide scene-flow/optical-flow/point-track annotations, and which provide only third-party preprocessed pseudo-labels. The sentence 'we use the data processing code or preprocessed data by DUSt3R, CUT3R, MonST3R, and CoTracker' is the only disclosure. For internet-image datasets such as MegaDepth, BlendedMVS, and DL3DV, dense GT depth and relative pose are not available, so the preprocessed data must be teacher estimates. Because LFv and LW are the principal signals that tie predicted scene flow to camera motion, teacher errors and static-scene bias in these pseudo-labels could directly suppress dynamic motion or incorrectly attribute it to ego-motion. The paper's own conclusion admits that 'the limited availability of scene flow data still constrains flow quality relative to depth,' underscoring that motion supervision is the weak link. This is not an internal inconsistency, but it means the empirical demonstration may not actually validate the clean, pose-regressor-free supervision story: the model's dynamic-scene capability could be dominated by a few synthetic datasets (Virtual KITTI 2, Spring, PointOdyssey, Dynamic Replica, Kubric, OmniWorld-Game) rather than by the proposed rigid-flow construction across all 22 datasets.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"Flow4R proposes a two-view transformer that predicts a per-pixel property set S = {P, F, W, C} (3D point positions, scene flow, pose weights, confidence) and uses closed-form operations (Eqs. 1-9) to derive camera pose, rigid/non-rigid flow decomposition, point tracks, focal length, and optical flow. The network is trained jointly on 22 static and dynamic datasets with losses for point positions (Eq. 10), 3D motion (Eq. 11), 2D motion (Eq. 12), pose weights (Eq. 13), and rigid flow (Eq. 14), and is evaluated on the WorldTrack benchmark for 3D point tracking and dynamic 3D reconstruction. The paper claims state-of-the-art performance and argues that scene flow is a unifying representation that removes the need for explicit pose regressors or bundle adjustment.","tokens_in":16542,"tokens_out":4127,"duration_ms":38436,"significance":"If the central claims hold, Flow4R offers an elegant and compact formulation: a single shared decoder predicts geometry and bidirectional motion, with downstream tasks obtained by simple algebraic operations. The derivations in §3.1 are internally consistent, and the evaluation against externally established benchmarks (WorldTrack) and baselines is a strength. The ablation in Table 3 supports the design choice of predicting P_vt rather than F. The model is comparatively small (0.4B parameters) and the qualitative results suggest meaningful generalization. However, the empirical demonstration is weakened by incomplete reporting of training-label provenance, and the 'state-of-the-art' claim is not supported on all benchmarks. These issues are fixable, but they are load-bearing for the paper's central supervision story.","major_comments":[{"comment":"The losses L_W and L_Fv require ground-truth depth and camera pose. Section 4 lists 22 datasets but never states which provide true GT depth/pose, which provide scene-flow/optical-flow/track labels, and which rely on pseudo-labels from DUSt3R, CUT3R, MonST3R, or CoTracker. For internet-image datasets such as MegaDepth, BlendedMVS, and DL3DV, dense GT depth and relative pose are not available, so their labels must be teacher estimates. Because L_W and L_Fv are the principal signals tying scene flow to camera motion, a large pseudo-label share would mean the claimed clean supervision is not actually demonstrated, and teacher static-scene bias could suppress dynamic motion. Please report the exact GT-vs-pseudo split per dataset and, ideally, an ablation or subset analysis with true GT only.","section":"§4 (Training Details) and §3.3 (Eqs. 13-14)"},{"comment":"The abstract and Section 1 claim 'state-of-the-art performance,' but Tables 1-2 show Flow4R trailing St4RTrack on Panoptic all-points (64.3 vs 69.7), on ADT dynamic points (70.9 vs 75.3), and on TUM-Dynamics reconstruction APD (79.87 vs 83.42). The claim should be qualified, e.g., 'competitive or state-of-the-art on most benchmarks.' The current phrasing overstates the results and is not supported by the data.","section":"Abstract, §1, Tables 1-2"},{"comment":"The pose-weight loss is described as 'self-supervised' or 'unsupervised,' but it uses ground-truth camera pose and depth through \\bar{T}\\bar{P}. When these come from preprocessed pseudo-labels, the loss is effectively teacher-supervised. This distinction matters for interpreting the learned pose weight map and for reproducibility. Please clarify which datasets provide the ground-truth pose used in Eq. (13) versus pseudo-pose.","section":"§3.3 (Eq. 13) and §4"}],"minor_comments":[{"comment":"The notation \\bar{P}_v is used before being defined. Define it explicitly as the ground-truth point position in the second view, obtained by applying the ground-truth camera pose to the back-projected depth.","section":"§3.3 (Eq. 14)"},{"comment":"The scale alignment step says 'align the predicted positions and flows of subsequent pairs by the scaling factors s1/s2, s1/s3...'. Clarify whether P_vt and W/C are also scaled, and specify how the average norm is computed when points are partially occluded or out of view.","section":"§3.2"},{"comment":"The focal-length estimation assumes a known optical center c and identical focal lengths on both axes. State whether c is assumed to be the image center or estimated, and discuss the impact of this assumption on the optical-flow computation in Eq. (9).","section":"§3.1 (Eq. 7)"},{"comment":"The symbol f is used both for focal length (Eq. 7) and optical flow (Eq. 9, Table 5). This is confusing; consider using f_x for focal length or \\mathbf{f} for optical flow.","section":"Notation"},{"comment":"The sentence 'we use the data processing code or preprocessed data by DUSt3R, CUT3R, MonST3R, and CoTracker' is vague. Please specify which datasets and which preprocessing pipelines are used, and cite exact versions or commit hashes if available.","section":"§4"},{"comment":"Several references are listed as 2026 preprints or in-press works (e.g., [2], [12], [16], [25], [42], [63], [75]). Verify that the cited versions are publicly available or update the citations.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The core formulation is sound and the paper is a solid candidate if the supervision-provenance issue is addressed. I do not see evidence of deliberate circularity — the concern is incomplete reporting. The SOTA claim should be calibrated against Tables 1-2. A table with per-dataset label types and an ablation or sensitivity analysis of pseudo-label impact would substantially strengthen the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — read the full text and checked the equations. My take: the core idea is genuinely good, and the paper deserves a real referee, but it needs revision before I'd trust the empirical claims at face value.\n\nWhat is new: the minimal per-pixel property set {P,F,W,C} with camera-space scene flow. It lets a single two-view transformer solve for geometry and motion without a pose regressor or bundle adjustment. The weighted SE(3) solve, the rigid/non-rigid decomposition, the focal-length and optical-flow projection, and the anchored-pair scale alignment are all clean and internally consistent. The idea of learning a pose-weight map without rigidity-mask supervision is a real contribution. The unified conversion of scene-flow, optical-flow, and point-track labels into camera-space flow supervision is also new and practically useful. The paper is honest about the F/P_vt choice, and ablates it (though not the pose-weight map).\n\nWhere it's soft: The abstract and intro say 'state-of-the-art' on 4D reconstruction and tracking, but the tables tell a more mixed story. Flow4R trails St4RTrack on Panoptic all-points (64.3 vs 69.7) and on TUM-Dynamics reconstruction (APD 79.87 vs 83.42, EPE 0.202 vs 0.185). It wins on ADT, DR, PO tracking, but 'state-of-the-art' needs qualification per-benchmark. No error bars, single runs, and no code/weights release. That limits how much we can lean on the reported margins.\n\nThe bigger concern is the training supervision. The paper lists 22 datasets but never says which contribute true GT depth/pose and which contribute pseudo-labels from DUSt3R/CUT3R/MonST3R/CoTracker. Since the rigid-flow loss and pose-weight loss both need depth and pose, a large fraction of the static 'rigid flow' supervision on internet-image data is likely teacher output. That doesn't sink the paper — the dynamic datasets (VKITTI2, Spring, PointOdyssey, Dynamic Replica, Kubric, OmniWorld-Game) provide real GT flow/tracks, and the core formulation is independent of label provenance — but it does mean the claimed 'unified motion supervision' is partly inherited teacher bias, and the authors should quantify the split and ideally ablate on a GT-only subset. The own conclusion concedes flow quality is data-limited, which fits.\n\nBottom line: worth reading, worth engaging, likely to be cited. Send to peer review with a request for code/weights, a per-dataset label-provenance table, a pose-weight ablation, and toned-down SOTA claims.","headline":"Genuinely new flow-centric formulation with clean math, but the unquantified GT-vs-pseudo supervision split and overstated SOTA claims need fixing.","tokens_in":17091,"tokens_out":3169,"would_cite":true,"duration_ms":29144,"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":"A single per-pixel scene-flow representation can reconstruct a 3D scene and track its motion from two images.","keywords":["scene flow","4D reconstruction","point tracking","camera pose estimation","dynamic scenes","two-view transformer","rigid flow supervision","video understanding"],"falsifier":"Train Flow4R on a balanced set of dynamic videos that have ground-truth depth and pose, then compare against the same model trained with the paper's synthesized rigid-flow labels for identical scenes. If the ground-truth-trained model's dynamic-point tracking accuracy is substantially higher, the rigid-flow supervision premise fails. A simpler check: run the model on a dynamic scene with no static pixels in view and see whether the solved camera pose remains accurate.","tokens_in":16011,"feed_emoji":"🔄","tokens_out":5846,"duration_ms":52930,"temperature":0.7,"pith_summary":"Flow4R sets out to prove that a single per-pixel quantity — camera-space scene flow — is enough to reconstruct a 3D scene and track its motion from just two images. Rather than predicting a shared-frame point map and a separate camera pose, the network outputs, for each pixel, a 3D point position, a scene flow vector, a pose weight, and a confidence score. Camera pose, rigid versus non-rigid motion, focal length, optical flow, and 3D point tracks are all derived from these outputs by closed-form weighted least-squares operations. Trained jointly on static and dynamic datasets, the model reaches leading or competitive accuracy on the WorldTrack tracking and reconstruction benchmarks with fewer parameters than several rivals. If right, this means one compact transformer can replace the usual separate pipelines for pose estimation, reconstruction, and tracking.","feed_headline":"Per-pixel scene flow unifies 4D reconstruction and tracking","feed_subtitle":"One two-view network outputs geometry, camera motion, object motion, and point tracks in a single pass.","key_machinery":"The central object is the camera-space scene flow map, defined per pixel as the vector carrying a 3D point from one image to its corresponding location in the other image's coordinate system and time. It is paired with a pose weight map that sums to one over pixels and encodes which pixels are reliable for solving the camera pose; the rigid transformation is the weighted least-squares SE(3) fit aligning predicted point positions with their flowed counterparts. This decomposition unifies the tasks: once the rigid transformation is known, rigid flow is computed directly, non-rigid flow is the residual, 3D point tracks are obtained by applying the inverse transformation, focal length by minimiz","core_discovery":"The paper's central claim is that 4D perception can be formulated entirely in terms of camera-space scene flow: the motion of each point relative to the camera, independent of any coordinate-system choice. The network takes two images and predicts the minimal property set consisting of point positions, scene flow, pose weights, and confidence maps for each image, symmetrically for both views. From these, the rigid transformation between views is solved by minimizing a pose-weight-weighted SE(3) error, splitting the scene flow into a rigid camera part and a non-rigid object part. 3D point positions in the other view's frame, focal length, and optical flow all follow by projection. The authors","pith_inferences":["The pose weight map is implicitly learning a prior over what 'static' means in the training data. A testable extension is to expose the model to videos with moving reference frames and check whether the map shifts accordingly; if it does not, the learned prior may be brittle outside its training distribution.","Since flow is defined relative to the camera, the representation is coordinate-system invariant; this suggests Flow4R-style models may generalize better to unseen camera trajectories than shared-frame point-map models, a claim that could be tested by cross-dataset pose-error evaluation.","The paper's own conclusion concedes that motion quality lags depth quality due to scarce flow data. A concrete next step is to add self-supervised cycle consistency on unlabeled video — forward flow from one frame to the next should invert backward flow — which could tighten dynamic tracking without new annotated data.","Deriving optical flow by projecting predicted 3D points gives a geometric constraint that 2D optical-flow losses do not directly enforce; comparing Flow4R's optical flow against a dedicated flow network on occluded regions would test whether this geometric prior helps."],"forward_implications":["Camera pose, focal length, optical flow, and 3D point tracks can be read out from a single set of per-pixel predictions, so one model can replace several task-specific modules.","Static reconstruction datasets with accurate depth can be used to supervise motion via synthesized rigid flow, so abundant static data helps dynamic tracking without needing dynamic ground truth.","Anchored pairing of frames lets the model extend to long video sequences with scale alignment, enabling tracking over many frames without bundle adjustment or pose-graph optimization.","Because the pose weight map learns which pixels are reliable, the model can switch reference frames at inference by changing this map, adapting to scenes where the intuitive 'static' reference differs."],"fun_headline_variants":["Scene flow unifies 4D reconstruction and tracking","One two-view network for geometry, motion, and point tracks","Flow4R: joint 4D reconstruction and tracking via scene flow","Single-pass scene flow solves 4D reconstruction and tracking"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The method depends on the correctness of the depth maps and camera poses used to synthesize its rigid-flow motion labels for static scenes; if a large fraction of those depths and poses are estimates produced by another model, the learned scene flow inherits those errors and a bias toward static scenes.","fun_headline_variants_meta":{"raw":{"variants":["Scene flow unifies 4D reconstruction and tracking","One two-view network for geometry, motion, and point tracks","Flow4R: joint 4D reconstruction and tracking via scene flow","Single-pass scene flow solves 4D reconstruction and tracking"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000281,"raw_usage":{"total_tokens":1484,"prompt_tokens":708,"completion_tokens":776,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":452,"completion_tokens_details":{"reasoning_tokens":706}},"tokens_in":452,"tokens_out":776,"duration_ms":7299,"temperature":1.0,"reasoning_tokens":706,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T23:21:04.864488+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train Flow4R on a balanced set of dynamic videos that have ground-truth depth and pose, then compare against the same model trained with the paper's synthesized rigid-flow labels for identical scenes. If the ground-truth-trained model's dynamic-point tracking accuracy is substantially higher, the rigid-flow supervision premise fails. A simpler check: run the model on a dynamic scene with no static pixels in view and see whether the solved camera pose remains accurate.","supporting_citations":[],"review_version":1}