{"id":"e63f3edc-c0f0-4627-a74c-821d6a7c517f","arxiv_id":"2504.13152","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A feed-forward network predicts two time-aware pointmaps per frame pair to simultaneously reconstruct dynamic scenes and track 3D points in a consistent world frame.","lead":"St4RTrack is a single feed-forward network that, given a pair of video frames, predicts two 3D point maps: one tracks the first frame's pixels to the second frame's time, and the other reconstructs the second frame's geometry, both in the first frame's world coordinates. This unifies dynamic 3D reconstruction and point tracking into one framework and enables test-time adaptation without 4D labels.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Eq. (1) two-frame mapping cannot determine where occluded or out-of-view points are at time j; Sec. 5 concedes this, so the world-frame tracking claim needs a visibility-restricted evaluation.","rationale":"The central claim is that fθ in Eq. (1) is a correct two-frame model of world-frame tracking. The load-bearing condition is observability: each output point must be determined by the two images. For occluded or out-of-view content this condition is violated by construction, not merely by training-data limitations. The paper's Sec. 5 limitation statement is an in-scope self-report that confirms exactly this failure mode. The reader identified the same assumption in the weakest_assumption field, and I agree with that characterization. The proposed split test would distinguish a benign interpretation (the method tracks only visible content accurately) from the full claim of dense long-range world-frame tracking. Because the paper already presents a new benchmark and promises code release, adding a visibility-masked analysis is a concrete and minimal condition for the claim to stand. The CONDITIONAL verdict is therefore appropriate and unchanged.","tokens_in":19183,"tokens_out":9689,"duration_ms":95572,"concrete_test":"Re-run the WorldTrack evaluation of Table 1 with visibility-masked splits. For Point Odyssey and Dynamic Replica, use the synthetic depth and occlusion maps to label each query point at each frame as visible or occluded; for ADT and Panoptic Studio, use the TAPVid-3D occlusion annotations. Report APD3D separately for (a) points visible in the paired frame Ij, (b) points occluded in Ij, and (c) pairs whose visible-overlap ratio with I1 is below a threshold (e.g., <30%) or whose frame gap is large (e.g., >24 frames). If APD3D on splits (b) and (c) is at or near chance while split (a) is high, the Eq. (1) two-frame assumption is falsified and the headline must be restricted to visible, overlapping-frame tracks.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. (1) defines fθ(Ii,Ij) = (iXi_j, iXj_j) with iXi_j a dense map of where every pixel of Ii is at time j. For this to be correct, the pair (Ii,Ij) must contain enough information to localize every such point at time j. This fails when a point is occluded in Ij, leaves the field of view, or is separated by a large camera motion or temporal gap: the image pair carries no evidence of that point's 3D position at time j, so the network can only hallucinate from priors. Inference in Sec. 3.1 always pairs I1 with each later frame, so the fraction of unobservable points grows with temporal gap and camera displacement. The paper's own Discussion (Sec. 5) states that 'scale misalignment, large camera movements, and occlusions are not fully resolved', which is precisely the regime where the Eq. (1) representation is not identifiable. If WorldTrack scores all points over 64 frames without visibility masking, the Table 1 APD3D numbers include these unobservable points, and the headline claim of long-range dense world-frame tracking is not established outside the visible/overlapping-frame regime. The ablations and TTA do not repair this, since the trajectory loss in Eqs. (5)-(7) is computed only where CoTracker provides pixels in frame j; occluded points receive no such signal.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes St4RTrack, a feed-forward framework for simultaneous dense 3D reconstruction and point tracking in a single world coordinate frame from monocular RGB video. For a pair of frames (Ii, Ij), the network predicts two pointmaps in the coordinate system defined by frame i, both evaluated at time j: the tracking pointmap iXi_j, which gives the 3D positions at time j of the content visible in frame i, and the reconstruction pointmap iXj_j, which gives the geometry of frame j. Chaining the first frame with every later frame yields long-range 3D point tracks and a per-frame reconstruction aligned to the first frame's coordinates. The model is initialized from MASt3R, pretrained on synthetic 4D datasets (Point Odyssey, Dynamic Replica, Kubric) with dense supervision, and refined at test time through a reprojection loss that uses CoTracker3 2D trajectories and MoGe monocular depth, with camera poses obtained via differentiable PnP. The authors introduce a new benchmark, WorldTrack, covering ADT, Panoptic Studio, Point Odyssey, and Dynamic Replica (50 sequences of 64 frames each), and report state-of-the-art results on world-frame 3D tracking (Table 1) and dynamic 3D reconstruction (Table 2), together with ablations of the test-time adaptation components.","tokens_in":19410,"tokens_out":11105,"duration_ms":100623,"significance":"If the results hold, this is a valuable contribution. The representation is minimal and clean: a small modification of the DUSt3R pointmap formulation unifies two traditionally separate tasks, and the proposed reprojection-based test-time adaptation provides a practical path toward self-supervised 4D training without dense 4D labels. The empirical gains over combinational baselines are consistent and large across multiple datasets, and the reconstruction branch also improves over MonST3R and DUSt3R. The new WorldTrack benchmark addresses a real gap in the literature, since prior benchmarks only evaluate tracking in camera coordinates. The architecture is simple, the pairwise inference is efficient (30 FPS on an RTX 4090 is claimed), and the authors plan to release code, models, and the benchmark. The paper is generally clearly written and the self-assessment in Section 5 is candid.","major_comments":[{"comment":"The APD3D metric as written in Eq. (12) is defined for a single threshold delta_3D and does not state whether it is evaluated over all query points or only those visible in each target frame. This is load-bearing because the two-frame mapping in Eq. (1) is underdetermined for points that are occluded or out of view at time j, and Section 5 explicitly concedes that occlusions and large camera movements are not fully resolved. If the reported scores include occluded or out-of-view points, the headline claim of long-range dense world-frame tracking is not established outside the visible/overlapping-frame regime; if the TAPVid-3D protocol does mask these points, that must be stated explicitly. I request that the authors specify the visibility handling in the evaluation and, ideally, report performance separately for visible and occluded/out-of-view points, along with a breakdown of the fraction of evaluated points that are occluded per dataset.","section":"Section 4.2 / Eq. (12), Table 1"},{"comment":"MonST3R is listed as a feed-forward tracking baseline in Table 1, but MonST3R is a dense reconstruction method that does not by itself output temporal correspondences. The paper never describes how the pointmap predictions of MonST3R are converted into the point tracks used to compute APD3D and EPE. Without this explanation, the 'MonST3R' row in the tracking tables is not reproducible and the reader cannot assess whether this non-tracking baseline is evaluated in a fair manner. Please provide a precise description of the procedure used to generate MonST3R tracks, including any matching or nearest-neighbor steps.","section":"Section 4.2 / Table 1"},{"comment":"The synthetic 4D supervision for the tracking branch is underspecified. The text states that scene mesh vertices provide 'sparse, masked supervision' for the tracking pointmap iX1_j, but it does not explain how world-space mesh vertices at time j are associated with pixels of frame 1, nor how the mask is defined. This supervision is the only source that teaches the network the time-dependent pointmap representation, so the training loss construction must be described in enough detail to be reproduced. Please clarify the exact loss terms, the masking, and how correspondences between frame-1 pixels and mesh vertices at other timestamps are obtained.","section":"Section 3.2 / Section 4.1"},{"comment":"The test-time adaptation protocol is described inconsistently. Section 4.1 states 'we run 500 optimization steps on a single sequence,' while Appendix C.1 specifies instance-level adaptation as training for 3 epochs on 300 sampled frames per epoch with batch size 4, which corresponds to a different number of gradient steps (75 steps per epoch, i.e., 225 total). The discrepancy between these descriptions must be resolved, and the main text should state the exact number of adaptation steps and the batch size used to produce the Table 1 results.","section":"Section 4.1 vs. Appendix C.1"}],"minor_comments":[{"comment":"The APD3D definition in Eq. (12) is a function of a single threshold delta_3D, but Table 1 reports one number per cell. Please state that the reported value is the average over the threshold set {0.1, 0.3, 0.5, 1.0} m and, if applicable, over time steps and query points; the caption phrase 'average points under distance' should be 'average percentage of points within distance.'","section":"Eq. (12), Table 1 caption"},{"comment":"MonST3R is described as a 'dynamic SLAM method' in the description of the second combinational baseline, but MonST3R is a feed-forward dense 3D reconstruction method, not an optimization-based SLAM system. Please correct this characterization to avoid confusion.","section":"Section 4.2, baselines paragraph"},{"comment":"The shorthand Xi_j := iXi_j is introduced after Eq. (3), but later equations and paragraphs do not consistently restate the left superscript. A short 'notational conventions' note listing the meaning of the superscript, left subscript, and right subscript would improve readability.","section":"Section 3.1, notation"},{"comment":"The ablation of the pretraining stage is only qualitative in the main text; Table 6 provides quantitative results but only on DR and PStudio and only for the test-time-adaptation variants. Consider reporting a quantitative feed-forward ablation on the full WorldTrack benchmark so the contribution of each component is measured on all four datasets.","section":"Section 4.5 / Table 6"},{"comment":"The rows labeled 'w/o trajectory loss', 'w/o depth loss', etc. should specify that they refer to the St4RTrack + TTA configuration with that loss component removed. As written, the baseline for these ablations is ambiguous.","section":"Table 6 caption"},{"comment":"The abstract claims 'dense 3D point tracks,' but the WorldTrack evaluation downsamples query points to approximately 1,000 per sequence. Please clarify whether the reported metrics are computed on the sampled queries or on the full dense pointmaps, and adjust the wording if the evaluation is on a sparse subset.","section":"Abstract / Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"This is a strong paper that likely merits acceptance after revision. The central representation is elegant and the empirical results are impressive, but the evaluation protocol needs tightening. The most important concerns are (1) the lack of explicit visibility handling in the APD3D metric, which directly affects the interpretation of the headline 'world-frame 3D tracking' claim, and (2) the unexplained construction of the MonST3R tracking baseline. Both are fixable with clarifications and additional reported breakdowns. The inconsistency between the main text and the appendix on adaptation steps is also easy to fix. I would ask the authors to address the visibility issue with a concrete response rather than a promise to add future experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"St4RTrack is worth your attention. The core idea is genuinely new: instead of predicting each frame's geometry in its own timestamp like MonST3R, the tracking branch predicts where the first frame's content is at time j, in the first frame's coordinates. That one change unifies dense 3D tracking and dynamic reconstruction in a single feed-forward network. The empirical results back it up: it beats combinational baselines on WorldTrack and dynamic reconstruction baselines on Point Odyssey and TUM-Dynamics, and the ablations show each test-time adaptation loss contributes.\n\nWhat's not to like? A few things, in proportion. The two-frame mapping in Eq. (1) cannot infer where occluded or out-of-view points are at time j—the image pair simply doesn't contain that information. The paper itself concedes this in Sec. 5. That's not a hidden flaw, but it means the \"long-range dense tracking\" claim is stronger than the representation can support in the all-points evaluation. If WorldTrack scores every query point over 64 frames without visibility masking, the APD3D numbers include points the network can only hallucinate. The fix is straightforward: report a visibility-restricted metric, or at least state how occluded points are handled.\n\nThe tracking comparison has a loose end: MonST3R is listed in the tracking table as a feed-forward baseline, but the paper never explains how tracks are extracted from a method that produces no temporal correspondence. The combinational baselines are described only briefly. And there are no error bars anywhere; given the benchmark filters sequences (PO filtering, TUM depth range), a reader can't tell if the 2-3 point APD gaps are noise.\n\nNone of this kills the paper. The representation is well-motivated, the TTA scheme is clever, and the authors are honest about the limitations. It deserves a serious referee. I'd send it to review, with requests for the code/model/benchmark release, a clear explanation of the MonST3R baseline, and either a visibility-restricted evaluation or a defended statement about why all-points scoring is appropriate.\n\nGood for a reading group on 4D perception.","headline":"A genuinely new pointmap redefinition for joint 4D tracking and reconstruction, with strong results and some honest evaluation gaps that need closing.","tokens_in":20031,"tokens_out":2879,"would_cite":true,"duration_ms":25619,"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":"St4RTrack claims that a single feed-forward network can simultaneously reconstruct dynamic scene geometry and track every visible point in a consistent world coordinate frame from monocular RGB video.","keywords":["4D reconstruction","long-range point tracking","world coordinate tracking","pointmap representation","monocular video","test-time adaptation","reprojection loss","dynamic scenes"],"falsifier":"A decisive check is a WorldTrack-style sequence with large camera translation between the anchor frame and a distant later frame plus a complete occlusion of a tracked region: the paper's own Discussion flags this regime as one where scale misalignment and occlusions are not fully resolved, so if the predicted world-frame tracks of the occluded points jump or drift when they reappear, the pairwise chaining of Eq. (1) has failed to carry correspondence through the gap. On the quantitative side, dynamic-point APD3D on the Point Odyssey split should not fall faster than the reconstruction branch's own error as the frame gap grows from 8 to 64; a sudden drop in tracking accuracy with gap size would expose the pairwise observability assumption as the load-bearing premise.","tokens_in":18937,"feed_emoji":"🎥","tokens_out":13066,"duration_ms":109806,"temperature":0.7,"pith_summary":"The paper argues that dynamic 3D reconstruction and dense 3D point tracking, usually treated as separate tasks, are two sides of the same coin once scene motion is accounted for, and it backs that claim with one network: St4RTrack takes a pair of frames and, in a single feed-forward pass, outputs both the second frame's geometry and where every pixel of the first frame has moved, expressed in one shared world coordinate frame. Repeating this pairing against a fixed first frame chains the predictions into long-range world-frame 3D tracks and per-frame reconstruction simultaneously. Because the predicted geometry also yields camera poses through a differentiable PnP step, the same architecture can adapt to new videos using only 2D trajectory and monocular depth pseudo-labels, with no 4D ground truth. The paper also introduces WorldTrack, a benchmark for world-frame 3D tracking, and reports that St4RTrack outperforms the combined baselines on both tracking and reconstruction.","feed_headline":"One pass reconstructs and tracks dynamic scenes in 3D from video","feed_subtitle":"A pair of pointmaps per frame pair yields dense world-frame 3D tracks and geometry from monocular RGB.","key_machinery":"The load-bearing object is the time-dependent pointmap, a per-pixel 3D map stamped with three factors: whose content it describes, which camera frame expresses it, and at which timestamp. St4RTrack realizes Eq. (1) with a DUSt3R-style siamese transformer whose two branches share weights but not roles: the tracking branch outputs $^{1}\\mathbf{X}^{1}_{j}$ (frame 1's content moved to frame $j$'s moment, in frame 1's world) and the reconstruction branch outputs $^{1}\\mathbf{X}^{j}_{j}$ (frame $j$'s geometry in the same world), with alternating self- and cross-attention letting the tracking branch borrow geometric cues from the reconstruction branch. Pairing the anchor frame with every later frame and chaining the outputs turns these pairwise predictions into a full sequence. The second mechanism is the adaptation loop: a differentiable Gauss-Newton PnP step recovers camera extrinsics from the reconstruction pointmap, then a scale-invariant reprojection loss against 2D tracks and monocular depth, plus a 3D self-consistency term between the two branches, fine-tunes the model on the target video without any 4D labels.","core_discovery":"The central claim is that the representation in Eq. (1) suffices: a learned function $f_\\theta(I_i, I_j) = (\\,^{i}\\mathbf{X}^{i}_{j},\\,^{i}\\mathbf{X}^{j}_{j})$ maps any pair of frames from a dynamic video into two pointmaps in one shared coordinate frame. Here the tracking pointmap $^{i}\\mathbf{X}^{i}_{j}$ records where the content visible in frame $i$ is located at time $j$, expressed in frame $i$'s coordinates, while the reconstruction pointmap $^{i}\\mathbf{X}^{j}_{j}$ records frame $j$'s own geometry in those same coordinates. Both pointmaps are thus at the same timestamp and in the same world, which is what makes geometry and correspondence dual again: once the network has accounted for motion, the dynamic problem falls back to the rigid case. Anchoring on the first frame, $\\{f(I_1,I_1), f(I_1,I_2), \\ldots, f(I_1,I_T)\\}$ yields the world-frame tracks $\\{^{1}\\mathbf{X}^{1}_{1}, \\ldots, ^{1}\\mathbf{X}^{1}_{T}\\}$ and the per-frame geometry $\\{^{1}\\mathbf{X}^{1}_{1}, ^{1}\\mathbf{X}^{2}_{2}, \\ldots, ^{1}\\mathbf{X}^{T}_{T}\\}$, and the paper reports that this unified prediction beats the best available combinations of separate trackers and reconstructors on WorldTrack, and improves dynamic reconstruction on Point Odyssey and TUM-Dynamics.","pith_inferences":["If the pairwise chaining is as robust as reported, adding temporal attention across frames, which the paper lists as future work, should push hardest on exactly the failure modes it names: scale misalignment, large camera movement, and occlusion, where the pairwise assumption is strained.","The scale-invariant reprojection losses inherit the accuracy of their pseudo-labels (CoTracker3 for 2D tracks, MoGe for depth); replacing those priors with stronger monocular estimators as they appear should improve the world-frame 3D tracks without any architectural change, a claim the paper does not make.","World-frame tracking separates camera motion from scene motion, so St4RTrack's output could seed optimization-based 4D reconstruction methods that currently optimize from scratch per video, potentially replacing their expensive initialization with a feed-forward prior.","Because the tracking pointmap is expressed in the anchor frame's coordinates by construction, the representation is naturally reversible: pairing the last frame as anchor would yield tracks backwards in time, which could serve video editing or re-timing applications that the paper does not discuss."],"forward_implications":["Dense 3D reconstruction and long-range point tracking become a single prediction, so improvements to the representation or the backbone improve both tasks at once rather than two separate pipelines.","Any monocular video, including footage with no ground-truth 3D or motion data, becomes usable training material, since the reprojection losses supply supervision from 2D tracks and monocular depth.","A consumer RGB video can yield a consistent world coordinate frame with per-frame camera poses, scene geometry, and dense 3D trajectories in one feed-forward pass, which is the input a downstream dynamic-view-synthesis or robotics system would need.","The WorldTrack benchmark gives the field a way to measure world-frame 3D tracking accuracy, and the paper's reported numbers on Panoptic Studio show the method also holds up when the camera is static.","During test-time adaptation only the tracking branch is fine-tuned; the paper reports that this preserves the view-alignment capability while correcting drifting points, making per-sequence adaptation converge in minutes."],"supporting_citations":[{"why":"DUSt3R contributes the two-pointmap representation and the siamese cross-attention architecture that St4RTrack adapts to the dynamic, time-indexed case.","marker":"[61]"},{"why":"MASt3R supplies the pretrained dual-branch transformer weights used to initialize St4RTrack before 4D synthetic pretraining.","marker":"[29]"},{"why":"MonST3R is the closest dynamic-scene prior, the formulation St4RTrack extends by adding temporal correspondence, and its feed-forward and globally aligned results are the main reconstruction baselines.","marker":"[67]"},{"why":"CoTracker3 provides the pseudo-ground-truth 2D trajectories that the scale-invariant reprojection loss aligns to during test-time adaptation.","marker":"[23]"},{"why":"MoGe provides the pseudo-ground-truth monocular depth used in the geometric-consistency term of the adaptation loss.","marker":"[59]"},{"why":"Point Odyssey supplies one of the three synthetic 4D datasets whose mesh vertices supervise the tracking branch's pointmap during pretraining.","marker":"[69]"},{"why":"Dynamic Replica, via the DynamicStereo dataset, supplies the second synthetic 4D training set with ground-truth geometry and motion.","marker":"[22]"},{"why":"EPro-PnP's derivative-based Gauss-Newton solver makes the camera-pose step differentiable so reprojection gradients flow back into both pointmaps.","marker":"[5]"},{"why":"SpatialTracker is the leading camera-frame 3D tracking method that the combinational baselines lift into world coordinates and that St4RTrack must beat.","marker":"[64]"},{"why":"TAPVid-3D supplies the APD evaluation protocol and the camera-coordinate sequences that WorldTrack re-targets to world coordinates.","marker":"[26]"}],"fun_headline_variants":["World-frame pointmaps sync geometry and tracks from video","Dual pointmaps unify 3D reconstruction and tracking in one shot","One model, two pointmaps: world-frame 3D tracking and geometry","Feed-forward frame pairs yield dense 3D tracks and geometry together","Simultaneous 4D world tracking and reconstruction from RGB video"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"From just the two images in a pair, the network must recover where every pixel of the first frame has moved by the second frame's moment, even when the camera has moved far, points are occluded, or the temporal gap is large, and the paper's own discussion concedes that scale misalignment, large camera movements, and occlusions are not fully resolved.","fun_headline_variants_meta":{"raw":{"variants":["World-frame pointmaps sync geometry and tracks from video","Dual pointmaps unify 3D reconstruction and tracking in one shot","One model, two pointmaps: world-frame 3D tracking and geometry","Feed-forward frame pairs yield dense 3D tracks and geometry together","Simultaneous 4D world tracking and reconstruction from RGB video"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000199,"raw_usage":{"total_tokens":1423,"prompt_tokens":1049,"completion_tokens":374,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":665,"completion_tokens_details":{"reasoning_tokens":282}},"tokens_in":665,"tokens_out":374,"duration_ms":4332,"temperature":1.0,"reasoning_tokens":282,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:13:58.000692+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive check is a WorldTrack-style sequence with large camera translation between the anchor frame and a distant later frame plus a complete occlusion of a tracked region: the paper's own Discussion flags this regime as one where scale misalignment and occlusions are not fully resolved, so if the predicted world-frame tracks of the occluded points jump or drift when they reappear, the pairwise chaining of Eq. (1) has failed to carry correspondence through the gap. On the quantitative side, dynamic-point APD3D on the Point Odyssey split should not fall faster than the reconstruction branch's own error as the frame gap grows from 8 to 64; a sudden drop in tracking accuracy with gap size would expose the pairwise observability assumption as the load-bearing premise.","supporting_citations":[{"cited_title":"Dust3r: Geometric 3d vision made easy","cited_arxiv_id":null,"evidence_quote":"DUSt3R contributes the two-pointmap representation and the siamese cross-attention architecture that St4RTrack adapts to the dynamic, time-indexed case."},{"cited_title":"MonST3R: A simple approach for estimating geometry in the presence of motion","cited_arxiv_id":null,"evidence_quote":"MonST3R is the closest dynamic-scene prior, the formulation St4RTrack extends by adding temporal correspondence, and its feed-forward and globally aligned results are the main reconstruction baselines."},{"cited_title":"Co- Tracker3: Simpler and better point tracking by pseudo- labelling real videos","cited_arxiv_id":null,"evidence_quote":"CoTracker3 provides the pseudo-ground-truth 2D trajectories that the scale-invariant reprojection loss aligns to during test-time adaptation."},{"cited_title":"Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision, 2024","cited_arxiv_id":null,"evidence_quote":"MoGe provides the pseudo-ground-truth monocular depth used in the geometric-consistency term of the adaptation loss."},{"cited_title":"Pointodyssey: A large-scale synthetic dataset for long-term point tracking","cited_arxiv_id":null,"evidence_quote":"Point Odyssey supplies one of the three synthetic 4D datasets whose mesh vertices supervise the tracking branch's pointmap during pretraining."},{"cited_title":"Dy- namicstereo: Consistent dynamic depth from stereo videos","cited_arxiv_id":null,"evidence_quote":"Dynamic Replica, via the DynamicStereo dataset, supplies the second synthetic 4D training set with ground-truth geometry and motion."},{"cited_title":"EPro-PnP: Generalized End-to-End Probabilistic Perspective-n-Points for Monocular Object Pose Estimation","cited_arxiv_id":"2303.12787","evidence_quote":"EPro-PnP's derivative-based Gauss-Newton solver makes the camera-pose step differentiable so reprojection gradients flow back into both pointmaps."},{"cited_title":"Spatialtracker: Tracking any 2d pixels in 3d space","cited_arxiv_id":null,"evidence_quote":"SpatialTracker is the leading camera-frame 3D tracking method that the combinational baselines lift into world coordinates and that St4RTrack must beat."},{"cited_title":"Tapvid-3d: A benchmark for tracking any point in 3d, 2024","cited_arxiv_id":null,"evidence_quote":"TAPVid-3D supplies the APD evaluation protocol and the camera-coordinate sequences that WorldTrack re-targets to world coordinates."}],"review_version":1}