{"id":"dfb5cda6-9e15-43bd-81f3-83f8ac092d7c","arxiv_id":"2507.17406","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"MoviCam adds a new real-world benchmark for moving-camera pose estimation with scene geometry, and PhysDynPose reports improved world-frame trajectory accuracy using scene-aware physics optimization.","lead":"MoviCam is a small, non-synthetic studio benchmark that records ground-truth camera trajectories, scene geometry, 3D human motion, and contact labels while a person moves around a room with a handheld camera. The paper also proposes PhysDynPose, a physics-based optimizer that combines an existing pose estimator and SLAM with scene height maps to recover human trajectories in world coordinates.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PhysDynPose's claimed balance is contradicted by its own Table 3: 68% non-flat scene penetration, worse than WHAM and GLAMR, even before considering GT contact/camera oracle inputs.","rationale":"The reader's verdict is REJECT, and my analysis supports that rejection, so I recommend UNCHANGED. I partially agree with the reader's weakest assumption: ground-truth contact labels are indeed a serious oracle input that compromises the evaluation. However, I would place the even more load-bearing concern on the paper's own Table 3: even with ground-truth contacts, scene mesh, and GT-initialized camera scale, PhysDynPose produces scene penetration in 68% of non-flat frames and is worse than both WHAM and GLAMR on that metric. This is not a speculative failure mode under predicted contacts; it is a direct contradiction of the 'physical plausibility' component of the central claim using the paper's own metrics. The global-trajectory advantages (W-MPJPE, RTE) are the strongest positive evidence, but they are exactly the metrics most affected by the GT camera-scale initialization and the root-supervision term in Eq. (4). Thus the paper needs either a non-oracle evaluation or a substantially better physical-plausibility result before the central claim can be accepted. No code or dataset release is available to independently verify the numbers, which further supports the REJECT verdict.","tokens_in":12558,"tokens_out":4531,"duration_ms":54165,"concrete_test":"Run a controlled ablation on MoviCam: feed PhysDynPose predicted contact labels (e.g., foot-to-mesh distances from 4DHumans output instead of ground-truth labels) and remove the ground-truth first-two-frame camera scale alignment, keeping everything else identical; if W-MPJPE/RTE advantages shrink or vanish, or if scene penetration stays above 50%, the reported global-trajectory and physical-plausibility claims are artifacts of privileged inputs.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that PhysDynPose achieves 'a balance between physical plausibility and motion accuracy, outperforming current approaches in reconstructing global trajectories'—is not supported by the paper's own evaluation. Two issues are load-bearing. First, the comparison is unfair in the method's favor: PhysDynPose receives ground-truth foot contact labels and the scene mesh as inputs (Section 4), and DROID-SLAM scale is aligned using the first two ground-truth camera frames (Section 4.1). These privileged inputs feed the friction-cone and no-sliding constraints in Eq. (5), directly improving the foot-sliding metric and the global trajectory metrics, so the reported wins over GLAMR, WHAM, and 4DHumans are not evidence for a deployable method. Second, even with those oracle inputs, Table 3 shows PhysDynPose still penetrates the scene in 68.13% of non-flat frames, with an average penetration depth of 119.23 mm—far worse than WHAM (16.19%, 40.23 mm) and GLAMR (1.74%, 2.92 mm)—and its average distance above the scene (377.37 mm) is worse than 4DHumans (285.54 mm). The 'physical plausibility' half of the claimed balance is therefore not established; the method only improves foot sliding. The conclusion's 'balance' claim is unsupported by the data as presented.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two contributions. First, it introduces MoviCam, a new benchmark captured in a controlled studio with a moving RGB camera, multi-view GT camera trajectories, scene geometry (mesh and height map), SMPL human poses, global trajectories, and foot-contact labels. Second, it proposes PhysDynPose, a physics-based optimization pipeline that combines 4DHumans kinematic estimates and DROID-SLAM camera trajectories with a PIP-style dual-PD physics optimizer extended by scene-height-map penetration checks and root supervision. The method is evaluated on MoviCam against GLAMR, WHAM, and 4DHumans, with reported improvements in global trajectory metrics (W-MPJPE, RTE) and foot sliding, while pose accuracy and scene-penetration metrics are mixed.","tokens_in":12848,"tokens_out":4068,"duration_ms":45694,"significance":"The MoviCam dataset is a potentially valuable resource: it is, to my knowledge, the first non-synthetic moving-RGB-camera dataset to provide scene geometry, global human motion, dynamic camera trajectories, and contact labels simultaneously. If the method claims were supported, the paper would make a useful step toward physics-based global human pose estimation in non-flat scenes. However, the evaluation as presented does not support the central claim that PhysDynPose robustly estimates human and camera poses from a single RGB camera while remaining physically plausible, because the method receives ground-truth contact labels and scene geometry as inputs and uses ground-truth camera frames to fix SLAM scale, and because its own penetration metrics are worse than two baselines. The dataset contribution may be salvageable, but the method contribution in its current form is not established.","major_comments":[{"comment":"The method description states that the inputs include 'scene mesh and foot contact labels' (Section 4, first paragraph), and Section 4.1 states that the DROID-SLAM trajectory is aligned using the first two frames of the ground-truth camera trajectory. The baselines in Section 5.2 are initialized only with ground-truth orientation and translation from the first two frames and do not receive scene mesh or contact labels. Consequently, the reported gains in W-MPJPE, RTE, and foot sliding in Tables 2 and 3 are partly attributable to test-time oracle information, not to the proposed optimization alone. In particular, the contact labels directly feed the no-sliding constraint in Eq. (5), making the foot-sliding improvement unsurprising. This undermines the abstract claim that the method 'robustly estimates both human and camera poses in world coordinates' from a single RGB camera.","section":"Section 4 and Section 4.1"},{"comment":"The physical-plausibility results in Table 3 (non-flat rows) contradict the conclusion that PhysDynPose 'achieves a balance between physical plausibility and motion accuracy.' PhysDynPose has 68.13% frames with scene penetration and 119.23 mm average penetration depth, whereas WHAM has 16.19% and 40.23 mm and GLAMR has 1.74% and 2.92 mm. Its average distance above the scene (377.37 mm) is also worse than 4DHumans (285.54 mm). Even with access to the ground-truth scene height map and contact labels, the method is less physically plausible than two baselines on the paper's own primary plausibility metrics. The only plausibility metric improved is foot sliding, which is directly supervised by the ground-truth contact labels.","section":"Table 3 and Section 7"},{"comment":"The ablation study does not support the claimed benefit of the full objective. On sequence 3, the 'Only Er' ablation achieves 51.25% penetration frames and 59.27 mm average penetration depth, while the full method 'Ours' achieves 87.12% and 137.57 mm. The full method improves W-MPJPE only slightly (490.61 vs 508.59) and matches RTE (0.70 vs 0.69). Thus the combined joint-angle and joint-position controller with contact constraints is substantially worse on scene penetration than the position-only controller, and the text in Section 5.4 that claims these components are complementary is not borne out by the penetration metrics.","section":"Tables 4 and 5"},{"comment":"The Limitations section acknowledges that PhysDynPose 'depends on ground truth camera initialization,' but it does not acknowledge the equally load-bearing dependence on ground-truth foot contact labels and the ground-truth scene mesh, even though these are listed as method inputs in Section 4. This omission matters because the paper's comparisons are framed as evidence of a deployable monocular RGB method, whereas the method actually requires scene geometry and contact supervision that are not available in the claimed application setting.","section":"Section 6"}],"minor_comments":[{"comment":"The ablation 'w/o height map & root supervision' removes two components simultaneously, so the individual contributions of the height map and root supervision to the reported degradation cannot be separated; an ablation removing one at a time would be more informative.","section":"Section 5.4"},{"comment":"References [17] and [18] cite the same SMPL paper in different formats; one duplicate should be removed or merged.","section":"References"},{"comment":"The table header 'Global Motion' is unclear because the entries include both real and simulated camera motion; a footnote defining the symbols would improve readability.","section":"Table 1"},{"comment":"The dataset contains 7 sequences and approximately 22,000 images, all captured in one studio scene; the paper should be cautious in describing this as 'complex scenes' and should state the scene diversity limits when positioning MoviCam as a general benchmark.","section":"Section 3.3"}],"recommendation":"reject","confidential_remarks":"The MoviCam dataset is a genuine contribution and could be the basis of a strong dataset-focused paper, or of a method paper that clearly separates oracle-aided evaluation from a fully automatic RGB-only evaluation. As submitted, however, the central method claim is not supported: the method uses ground-truth contact labels, scene geometry, and ground-truth camera frames for scale alignment, and its own penetration metrics are worse than baselines. These are load-bearing issues that affect the paper's main conclusion, not merely presentation problems."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, MoviCam is the genuinely new part: a non-synthetic benchmark with a moving camera, registered scene mesh, world-frame SMPL motion, and foot contact labels. That combination does not exist elsewhere, and the capture pipeline (checkerboard hand-eye calibration, 120-camera studio) is careful and plausible. Second, the method evaluation does not support the paper's headline claim of a 'balance between physical plausibility and motion accuracy.' The method takes ground-truth contact labels and the scene mesh as inputs, and DROID-SLAM scale is aligned to the first two ground-truth camera frames. Those oracle inputs feed directly into the friction and no-sliding constraints, so the improved W-MPJPE, RTE, and foot-sliding numbers are not evidence for a deployable system. Even with that help, Table 3 shows the method still penetrates the scene in 68% of non-flat frames—far worse than WHAM (16%) and GLAMR (1.7%)—and its MPJPE is worse than 4DHumans. The conclusion overstates what the data show.\n\nWhat the paper does well is honest reporting of limitations and a clear statement of what the dataset provides. The related-work comparison is thorough, and the discussion of why existing datasets fall short (no scene geometry, simulated camera motion, missing contact labels) is accurate. The ablation is single-sequence, with no error bars, so it can only be indicative.\n\nThe soft spots are real but not equal in weight. The oracle inputs are the load-bearing problem; the authors disclose them, but the conclusions do not account for them. A second issue is the physical-plausibility metric: the method is worse than baselines on penetration, and the text explains this away by arguing the baselines are 'positioned unrealistically high.' That may be true, but then the penetration metric alone is misleading, and the paper should present a composite view or justify why height error should trade off against penetration.\n\nWho is this for? Anyone working on global human pose estimation with moving cameras or on human-scene interaction benchmarks. The dataset deserved a serious referee, and the paper deserves a chance to be improved. My recommendation: major revision. Either remove the oracle inputs from the method evaluation and show what happens with predicted contacts and no GT scale, or restructure the paper to present MoviCam as the primary contribution with PhysDynPose as an illustrative baseline. The current form conflates the two, and the method claims do not survive contact with the evidence.","headline":"The MoviCam dataset is a real contribution that fills a gap; the PhysDynPose method is an incremental integration whose reported gains are undermined by oracle inputs and its own physical-plausibility metrics.","tokens_in":13429,"tokens_out":1323,"would_cite":true,"duration_ms":16607,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Replacing the flat-ground assumption with a real scene height map in a physics optimizer yields more accurate global human motion from a moving camera.","keywords":["3D human pose estimation","physics-based optimization","moving camera","global trajectory","scene geometry","foot contact labels","monocular RGB video","benchmark dataset"],"falsifier":"Run the method on the same sequences but replace the ground-truth contact labels with labels predicted by an automatic contact estimator, leaving everything else fixed; if the reported advantages over the kinematic baseline vanish or reverse, the headline result depends on oracle contact information.","tokens_in":12316,"feed_emoji":"🎥","tokens_out":8662,"duration_ms":79963,"temperature":0.7,"pith_summary":"This paper contends that the flat-ground assumption built into most monocular and physics-based human pose trackers is the reason they fail when the camera moves or the terrain is uneven. To make that case, the authors introduce MoviCam, a real (non-synthetic) dataset whose ground truth includes camera trajectories, scene meshes, 3D human motion, and foot-contact labels. They then propose PhysDynPose, which takes kinematic estimates from a monocular pose estimator and a SLAM camera trajectory and refines them with a physics optimizer that consults the scene's height map. On this benchmark they report that prior methods suffer from drift, penetration, and floating, while PhysDynPose produces the most accurate global trajectories and the least foot sliding.","feed_headline":"Scene-aware physics optimizer fixes moving-camera pose drift","feed_subtitle":"New MoviCam benchmark and PhysDynPose method recover global human motion on non-flat terrain from a single moving camera.","key_machinery":"The central mechanism is a physics optimizer whose character is simulated as a floating-base humanoid with joint torques and contact forces constrained by the equation of motion, a friction cone, and a no-sliding condition. Two modifications carry the claim: the penetration test for a contacting foot uses the scene height map $h(x,z)$ rather than a flat floor, and the root translation is supervised using future frames to prevent long-sequence drift. The dataset's ground-truth camera motion, mesh, and contact labels make these effects measurable.","core_discovery":"PhysDynPose shows that incorporating a scene height map into a physics-based pose optimizer, in place of the usual flat-ground inequality, and adding root-position supervision from future frames yields global human trajectories that are more accurate and less prone to drift than prior kinematic and physics-based methods on moving-camera, non-flat scenes. The paper further establishes MoviCam as a benchmark for this setting, and reports it is the first non-synthetic dataset to combine ground-truth camera trajectories, scene geometry, 3D human motion, and foot-contact labels.","pith_inferences":["A natural stress test is to replace the ground-truth contact labels with automatic ones; the paper's reported gains over the kinematic baseline would likely shrink, since contact is an explicit input to the optimizer.","The same scene-mesh and height-map machinery could be extended to contacts beyond the feet, such as hands or torso, once the optimizer monitors additional body parts, which would broaden it to climbing and crouching interactions.","Because MoviCam provides accurate camera ground truth alongside RGB video, it could serve as a testbed for monocular SLAM scale calibration in human-centric scenes, where scale ambiguity is usually unresolved.","The manual PD-controller gains could be learned from data, which would likely make the optimizer more robust to different body shapes and interaction styles than the fixed settings used here."],"forward_implications":["Because the physics module is plug-and-play and needs no training, it can be attached to any kinematic pose estimator and any camera trajectory estimator, provided scene geometry and contact information are available.","The height-map penetration check can be reused by other physics-based trackers, replacing the flat-ground inequality and extending them to non-planar scenes.","Root supervision with future frames is a simple addition that reduces long-sequence drift, and the paper shows removing it degrades global trajectory error.","MoviCam supplies ground truth for metrics that go beyond per-joint error, such as scene penetration, elevation above the scene, jitter, and foot sliding, enabling comparisons in world coordinates."],"supporting_citations":[{"why":"It supplies the dual-PD controller physics optimizer, with the equation of motion, friction cone, and no-sliding constraints that the proposed method extends.","marker":"[39]"},{"why":"It supplies the monocular kinematic human pose and shape estimates that initialize the physics refinement.","marker":"[5]"},{"why":"It provides the camera trajectory estimates used to lift the kinematic pose into the world frame.","marker":"[30]"},{"why":"It establishes the physics-based pose optimization approach and the foot-scene contact labeling rule used for ground truth.","marker":"[24]"},{"why":"It represents the prior dynamic-camera dataset with global human and camera trajectories but no scene geometry, which MoviCam complements.","marker":"[12]"},{"why":"It represents the prior scene-aware urban dataset that lacks foot-contact labels, motivating the new contact annotations.","marker":"[4]"},{"why":"It supplies the markerless multi-view motion capture system used to obtain ground-truth human motion.","marker":"[31]"},{"why":"It supplies the physics simulator used for humanoid dynamics, height-map raycasting, and contact checking.","marker":"[3]"}],"fun_headline_variants":["Physics-based pose tracking adapts to moving cameras and uneven ground","MoviCam and PhysDynPose tackle non-flat moving-camera scenes","Scene-aware optimizer improves human pose under camera motion","New dataset and method for robust human motion from moving RGB","PhysDynPose reduces drift by using scene geometry in pose optimization"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The optimizer receives ground-truth foot-contact labels and a pre-scanned scene mesh as inputs; if those must be predicted from the RGB video, the physics constraints become noisy and the reported trajectory and foot-sliding gains may not hold.","fun_headline_variants_meta":{"raw":{"variants":["Physics-based pose tracking adapts to moving cameras and uneven ground","MoviCam and PhysDynPose tackle non-flat moving-camera scenes","Scene-aware optimizer improves human pose under camera motion","New dataset and method for robust human motion from moving RGB","PhysDynPose reduces drift by using scene geometry in pose optimization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000164,"raw_usage":{"total_tokens":1225,"prompt_tokens":901,"completion_tokens":324,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":235}},"tokens_in":517,"tokens_out":324,"duration_ms":4757,"temperature":1.0,"reasoning_tokens":235,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:48:35.662483+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the method on the same sequences but replace the ground-truth contact labels with labels predicted by an automatic contact estimator, leaving everything else fixed; if the reported advantages over the kinematic baseline vanish or reverse, the headline result depends on oracle contact information.","supporting_citations":[{"cited_title":"Pybullet, a python mod- ule for physics simulation for games, robotics and machine learning, 2016","cited_arxiv_id":null,"evidence_quote":"It supplies the physics simulator used for humanoid dynamics, height-map raycasting, and contact checking."},{"cited_title":"Physi- cal inertial poser (pip): Physics-aware real-time human mo- tion tracking from sparse inertial sensors","cited_arxiv_id":null,"evidence_quote":"It supplies the dual-PD controller physics optimizer, with the equation of motion, friction cone, and no-sliding constraints that the proposed method extends."},{"cited_title":"Humans in 4d: Re- constructing and tracking humans with transformers","cited_arxiv_id":null,"evidence_quote":"It supplies the monocular kinematic human pose and shape estimates that initialize the physics refinement."},{"cited_title":"Droid-slam: deep visual slam for monocular, stereo, and rgb-d cameras","cited_arxiv_id":null,"evidence_quote":"It provides the camera trajectory estimates used to lift the kinematic pose into the world frame."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It establishes the physics-based pose optimization approach and the foot-scene contact labeling rule used for ground truth."},{"cited_title":"Emdb: The electromagnetic database of global 3d human pose and shape in the wild","cited_arxiv_id":null,"evidence_quote":"It represents the prior dynamic-camera dataset with global human and camera trajectories but no scene geometry, which MoviCam complements."},{"cited_title":"Sloper4d: A scene-aware dataset for global 4d human pose estimation in urban environments","cited_arxiv_id":null,"evidence_quote":"It represents the prior scene-aware urban dataset that lacks foot-contact labels, motivating the new contact annotations."},{"cited_title":"The captury, 2020","cited_arxiv_id":null,"evidence_quote":"It supplies the markerless multi-view motion capture system used to obtain ground-truth human motion."}],"review_version":1}