{"id":"caf666e8-69e6-44b9-ae37-f0ca1f479ebd","arxiv_id":"1908.11676","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A markerless motion capture pipeline that estimates global 3D ski poses from rotating cameras by tying unknown camera orientation to tracked background motion during bundle adjustment.","lead":"This paper describes a multi-camera method that reconstructs a skier's global 3D position and body pose from ordinary videos, even when the cameras rotate freely. The system tracks background features to infer the unknown camera orientations during a joint 3D pose optimization.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Uncalibrated accuracy is contingent on ORB homography rotation estimates; the paper's own SfM failure note and the Ours-B collapse make this the load-bearing condition.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing point: the uncalibrated method's accuracy depends on homography rotation differences from background features. I read the paper in good faith and found no internal inconsistency in the main pipeline; the calibrated results and the multi-camera ablations in the appendix support the local-pose claims. The decisive evidence is the Ours-B baseline: without the rotation-difference term, uncalibrated global MPJPE jumps from 0.701 m to 7.59 m, so the central uncalibrated claim is carried by the ORB homography cue. The paper itself concedes that structure-from-motion fails on the same footage for lack of discriminative patterns, which makes the absence of any low-texture or feature-count sensitivity test a genuine gap. The concern does not refute the paper; it conditions the claim on scene content. Since the reader's verdict is already CONDITIONAL, no adjustment is needed. I considered the questionable robust norm in Eq. 8 (the claimed 'negligible influence' of outliers is not what the formula does), but it is secondary: the uncalibrated experiments would still be dominated by the rotation-cue dependency, and the empirical comparisons are not built on that norm's theoretical justification.","tokens_in":17707,"tokens_out":8600,"duration_ms":84609,"concrete_test":"Re-run the uncalibrated MV-Ski evaluation (Table 2) with ground-truth relative rotations substituted for the ORB-derived ΔR in Eq. 14, keeping all other components identical. If global MPJPE and CoM error stay near 0.701 m and 0.78 m, the reported uncalibrated error is not caused by ORB homography noise, and the concern does not land. If they drop substantially toward the calibrated 0.092 m, the ORB ΔR estimates are the bottleneck; a follow-up sweep that degrades background texture (e.g., replacing the region outside the athlete crop with uniform snow) would then confirm the deployment risk.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The uncalibrated branch of the central claim rests on the homography rotation differences ΔR^{f,c} computed from ORB features in the static background (Section 4.5, Eqs. 12-13). These enter the final objective through E_rot with λ_rot=10000 (Eq. 14), so errors in ΔR are strongly amplified rather than downweighted. The Introduction states that off-the-shelf structure-from-motion methods 'fail on skiing footage due to the large zoom, large distance and view angle between cameras, and lack of discriminative patterns on the ski slope' — the same footage used for evaluation. Yet the paper provides no sensitivity analysis: no variation of background texture, feature count, occlusion, or moving background content, and the only evidence for ΔR accuracy is one qualitative plot (Fig. 5). The ablation Ours-B, which optimizes rotations without Eq. 14, collapses to 7.59 m global MPJPE; the reported 0.701 m uncalibrated accuracy is therefore not a property of the joint pose+rotation optimization itself, but of the background-feature conditioning. If a deployment scene has a textureless snowfield, moving spectators, or occluded background, the measured ΔR will be noisy or wrong and the central claim 'accurate global 3D poses from images only' has no support.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a multi-camera motion capture system for alpine skiing that works with pan-tilt cameras whose orientation is unknown. The pipeline first detects a tight bounding box around the athlete and estimates 2D joint positions with a fine-tuned OpenPose network, using a newly introduced SV-Ski dataset for training. The 3D pose is then recovered by a bundle adjustment that parameterizes both the pose and the camera rotations in a discrete-cosine basis. For the uncalibrated case, the camera rotations are initialized with a look-at scheme and then jointly optimized with the pose under a consistency term that ties inter-frame rotations to homographies computed from ORB features in the static background. On the MV-Ski dataset, the method reports a global MPJPE of 0.092 m in the calibrated case and 0.701 m in the uncalibrated case, with centered errors of 0.077 m and 0.090 m respectively; an ablation without the homography consistency term (Ours-B) degrades to a global MPJPE of 7.59 m. The paper also contributes a new single-view alpine skiing pose dataset.","tokens_in":18002,"tokens_out":8169,"duration_ms":72370,"significance":"If fully validated, the paper addresses a practical and important problem: obtaining global 3D pose and speed measurements for athletes from multiple rotating cameras without cumbersome orientation calibration. The core idea of conditioning camera-rotation optimization on homographies derived from static background features is a sensible alternative to full structure-from-motion, and the DCT parametrization of pose and rotation is a clean way to enforce smoothness. The new SV-Ski dataset is a useful contribution to the community. However, the central claim is currently overstated relative to the actual input requirements, the robust-norm formulation contains a mathematical error, and the main comparison baseline is unpublished. These issues need to be resolved before the performance figures can be taken as fully supporting the stated claims.","major_comments":[{"comment":"The robust reweighting function g(x,y) = (n(0) - n(e_rep)) e_rep is not robust to outliers. Since n(e) tends to 0 as e grows, the prefactor (n(0) - n(e)) saturates to the constant n(0), so the whole expression grows linearly with the residual e_rep rather than downweighting large residuals. The sentence 'Using this norm with σ²=100, outlier points have negligible influence on the energy function' is therefore incorrect. Please replace Eq. (8) with an actual robust norm (for example, a truncated quadratic or a Geman-McClure penalty) or remove the robustness claim; as written, the optimization's tolerance to 2D detection errors is not supported.","section":"§4.4, Eq. (8)"},{"comment":"The abstract's claim that the method estimates accurate global 3D poses 'from images only' conflicts with the assumptions stated in Section 4: known camera intrinsics, known relative camera positions, and limb lengths ℓ(i,j) that are taken from the ground truth data. Using test-subject ground-truth limb lengths in Eq. (10) constitutes evaluation leakage and makes the reported numbers more optimistic than a strict 'from images only' reading implies. The claims should be revised to state the actual required inputs, and the limb-length supervision should be either removed or explicitly justified as a manual measurement that would be available in practice.","section":"Abstract and §4.4, Eq. (10)"},{"comment":"The main comparison baseline, 'Ostrek et al. (unpublished data)', is not publicly available or citable. This makes the claim that the proposed method performs comparably to this baseline unverifiable. Please replace it with a published, accessible baseline or release the necessary implementation and experimental details so that the comparison can be reproduced.","section":"§5.2, Table 2"},{"comment":"The uncalibrated accuracy is critically dependent on the homography-derived rotation differences ΔR computed from ORB features in the static background. The introduction reports that off-the-shelf structure-from-motion fails on the same skiing footage 'due to the large zoom, large distance and view angle between cameras, and lack of discriminative patterns on the ski slope,' but the only validation of ΔR is a single qualitative plot (Fig. 5). Because the ablation Ours-B, which does not use Eq. (14), collapses to a global MPJPE of 7.59 m, the 0.701 m uncalibrated result is contingent on the background-feature conditioning. Please provide a quantitative evaluation of the ΔR estimates across cameras and sequences and a sensitivity analysis with respect to background texture, feature count, and moving objects, or explicitly restrict the claim to scenes with sufficiently textured static backgrounds.","section":"§4.5, Eqs. (12)-(14)"}],"minor_comments":[{"comment":"There is a typographical error in the Global speed MAE row: '1.87.±2.64' should presumably read '1.87±2.64'.","section":"Table 2"},{"comment":"In configuration (b), the MV-Ski all-joints MPJPE of 0.0137±0.1236 appears inconsistent with the corresponding PCK of 65.51; this is a much lower error than the SV-Ski all-joints value of 0.0807 despite a much lower PCK. Please verify these numbers.","section":"Table 1"},{"comment":"Figure 5 shows estimated rotation differences for only one camera and one sequence; consider adding a summary of errors across all cameras and test sequences, or error bars.","section":"§5, Fig. 5"},{"comment":"The DCT basis sizes N_Π and N_Γ are chosen empirically (25/11 and 11), and the paper does not report how sensitive the results are to these values. A small sensitivity analysis would help the reader understand the robustness of the parametrization.","section":"§4.3, §4.5"},{"comment":"Please clarify whether the limb lengths in Eq. (10) are measured from the same test subjects whose ground-truth poses are used for evaluation; if so, this should be explicitly stated as a limitation in the experiments section.","section":"§4.4, Eq. (10)"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of a computer vision venue and addresses a real application in sports biomechanics. My main concerns are the overstated 'from images only' claim, the incorrect robust-norm formulation in Eq. (8), the use of an unpublished baseline, and the lack of sensitivity analysis around the background-feature conditioning. These are all addressable in a major revision, so I do not recommend rejection, but the current version should not be accepted as is."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know before reading. First, this is a solid engineering paper, not a conceptual breakthrough: it combines known pieces—DCT parametrization, bundle adjustment, homography-derived rotation priors—into a pipeline that works on a genuinely hard skiing dataset, and it releases a new 1982-frame single-view ski annotation set. Second, the headline claim \"accurate global 3D poses from images only\" is true only under a condition the paper never probes: the background must have enough static, matchable features. The uncalibrated accuracy of 0.70 m global MPJPE is achieved with a rotation penalty weighted at lambda_rot = 10000; without it, the ablation baseline collapses to 7.59 m. So the deployment value rests almost entirely on the ORB homography estimates being right.\n\nWhat is genuinely new and good: the DCT basis for both pose and camera Euler angles is a clean smoothness prior, and using background homographies to condition PTZ rotation estimation is a natural fit for sports footage. The calibration-pole augmentation and the SV-Ski dataset are practical, useful contributions, and the paper is transparent about the effect of adding one MV-Ski training sequence. The error decomposition into global, centered, and normalized MPJPE is the right way to show that local pose quality survives even when global translation drifts.\n\nSoft spots, in rough order of seriousness. (1) The robust norm in Eq. 8 is not robust in the way claimed: g(x,y) = (n(0)-n(e))e grows linearly for large e, so outliers still have unbounded influence. The statement that outlier points have negligible influence is wrong. (2) The main comparison baseline, Ostrek et al., is unpublished, so the central quantitative context is not independently verifiable. (3) The 2D detector is fine-tuned on one MV-Ski training sequence from the same dataset used for evaluation; the paper is honest about this, but it still weakens any generalization claim. (4) The stress-test concern is real: there is no sensitivity analysis varying background texture, feature count, occlusion, or moving distractors. The paper notes that SfM fails on this slope due to lack of discriminative patterns, so the ORB homographies clearly worked in this particular footage, but the only evidence is one qualitative plot and no failure statistics. That is the main gap if the paper wants to claim broad real-world applicability.\n\nI would send this to peer review. The core setup is plausible, the engineering is honest, and the dataset is useful. The revision should add a sensitivity analysis for the homography rotation term and correct or clarify the robust-norm claim.","headline":"A practical sports-mocap pipeline that couples DCT bundle adjustment with ORB homography rotation priors; the central claim holds in the tested setting, but the uncalibrated branch leans on a cue the paper never stress-tests.","tokens_in":18538,"tokens_out":2360,"would_cite":true,"duration_ms":22831,"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":"Joint optimization of athlete pose and pan-tilt camera rotation, guided by background feature motion, recovers accurate global 3D poses when camera orientations are unknown.","keywords":["markerless motion capture","pan-tilt cameras","unknown camera orientation","bundle adjustment","homography","3D human pose estimation","alpine skiing","background feature tracking"],"falsifier":"Run the pipeline on a pan-tilt sequence whose background is textureless, such as a uniform snow slope or blank wall, or has moving people behind the athlete, keep the same optimizer and weights, and compare the global MPJPE against the calibrated result; if the rotation penalty no longer keeps the joint optimization near the calibrated pose, the conditioning claim fails. A simpler version: remove the background-feature rotation term and confirm that the global error jumps toward the reported 7.59 m baseline.","tokens_in":17505,"feed_emoji":"🎿","tokens_out":9638,"duration_ms":79934,"temperature":0.7,"pith_summary":"Alpine ski coaches need the skier's global position, speed, and joint angles, but body-worn sensors are cumbersome and manual annotation is slow, while monocular pose methods cannot recover global scale or position. This paper tries to establish that a multiview markerless system can deliver those global measurements from plain video even when the cameras pan and tilt with unknown orientation. The core proposal is a bundle adjustment that optimizes the 3D pose and the per-frame camera rotations together, while adding a term that ties the rotations to relative motion measured from tracked background features. When camera orientations are estimated this way, local pose accuracy stays close to the fully calibrated case, while the global position error is about 0.7 meters over a large capture volume. The payoff is a practical performance-analysis tool for outdoor sports where fixed-camera calibration is infeasible.","feed_headline":"Uncalibrated pan-tilt cameras still capture 3D ski pose","feed_subtitle":"Background motion cues replace camera calibration, giving coaches global position and speed from plain video.","key_machinery":"The central object is the per-frame camera rotation extracted from background homographies, where a homography is the planar projective mapping between two views. For a pan-tilt camera with fixed position, points in consecutive frames $f$ and $f+1$ are related by $H^{f,c} = K^{f+1,c}\\,\\Delta R^{f,c}\\,(K^{f,c})^{-1}$, so the relative rotation $\\Delta R^{f,c}$ can be read off from a homography computed from tracked features outside the athlete's bounding box. These measured rotations enter the bundle adjustment through a penalty term that compares them with the rotations produced by the jointly optimized pose, weighted by $\\lambda_{\\text{rot}} = 10000$. A complementary piece is the discrete cosine basis parametrization of both the pose and the camera angles, which keeps the trajectories smooth and reduces the number of free parameters in the joint optimization.","core_discovery":"The paper's central claim is that the orientation of pan-tilt cameras can be recovered jointly with an athlete's global 3D pose, provided the optimization is conditioned on background feature motion. For a camera fixed in space, consecutive frames are related by a homography that encodes the inter-frame rotation; the paper detects features in the static background outside the athlete's bounding box, computes per-frame rotations from those features, and adds a penalty that keeps the optimized camera rotations consistent with them. With this constraint, the bundle adjustment no longer diverges, and the paper reports, with six cameras and unknown rotations, a global mean per-joint error of $0.701 \\pm 0.219$ m, a centered error of $0.090 \\pm 0.085$ m, and biomechanical angles within a few degrees of the calibrated result. The same pipeline without the rotation constraint yields a global error of $7.59$ m, and the paper presents the method as the piece that makes uncalibrated multi-camera motion capture practical for fast outdoor sports.","pith_inferences":["If the conditioning on background homographies is the real cause of the convergence gain, the same mechanism should transfer to other outdoor sports with static scenery, such as cycling or running courses, where fixed camera calibration is equally impractical.","The rotation penalty's very large weight suggests that homography noise will translate directly into global drift; a natural extension would be to fuse the same background-feature rotations with cheap inertial angular measurements from the camera to correct low-frequency drift.","A testable extension would replace the hand-crafted background features with a learned dense feature matcher, which could extend the method to low-texture or heavily blurred backgrounds, the regime where the paper's own structure-from-motion baseline fails."],"forward_implications":["Coaches can obtain global center-of-mass position, speed, and joint angles from multi-view video of a fast outdoor sport without camera-orientation calibration, using only relative camera positions and once-calibrated intrinsics.","The gap between calibrated and uncalibrated performance is concentrated in global position: centered joint errors and biomechanical angles stay close to the calibrated numbers, so the method is usable for technique analysis even where absolute position matters less.","Larger camera counts help the uncalibrated case most on global metrics, with the biggest improvement coming when a third camera is added.","Monocular methods that recover only relative pose cannot supply global speed; this approach does, with a reported speed error of about 1.9 m/s against skiers moving at 15–20 m/s."],"supporting_citations":[{"why":"supplies the inverse discrete cosine transform parametrization that keeps pose and camera rotation trajectories smooth during bundle adjustment.","marker":"[4]"},{"why":"supplies the 2D joint detections from image crops that the multi-view bundle adjustment reprojects.","marker":"[6]"},{"why":"provides the MV-Ski benchmark data and the global center-of-mass and biomechanical references used for evaluation.","marker":"[11]"},{"why":"establishes the homography relation between consecutive frames that turns background motion into relative camera rotations.","marker":"[16]"},{"why":"detects a tight bounding box around the athlete, whose interior is excluded so that only static background features are tracked.","marker":"[22]"},{"why":"provides the semi-supervised monocular baseline and multi-view dataset that the paper compares against.","marker":"[35]"},{"why":"finds and describes the background keypoints used to estimate each camera's inter-frame homography.","marker":"[37]"},{"why":"is the closest prior method that jointly optimizes human pose and uncalibrated camera rotations, which this paper's background conditioning stabilizes.","marker":"[40]"}],"fun_headline_variants":["Background flow replaces calibration for 3D ski pose","Uncalibrated PTZ cameras track skier 3D pose via background motion","Joint camera-pose optimization sidesteps calibration for ski tracking","Background motion cues replace camera calibration for 3D ski pose","Pan-tilt cameras need no calibration for 3D ski tracking"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the background outside the athlete's bounding box is static and textured enough for tracked features to yield accurate per-frame camera rotations, with camera positions and intrinsic parameters known in advance.","fun_headline_variants_meta":{"raw":{"variants":["Background flow replaces calibration for 3D ski pose","Uncalibrated PTZ cameras track skier 3D pose via background motion","Joint camera-pose optimization sidesteps calibration for ski tracking","Background motion cues replace camera calibration for 3D ski pose","Pan-tilt cameras need no calibration for 3D ski tracking"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001733,"raw_usage":{"total_tokens":6850,"prompt_tokens":947,"completion_tokens":5903,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":5813}},"tokens_in":563,"tokens_out":5903,"duration_ms":39219,"temperature":1.0,"reasoning_tokens":5813,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:08:55.462411+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the pipeline on a pan-tilt sequence whose background is textureless, such as a uniform snow slope or blank wall, or has moving people behind the athlete, keep the same optimizer and weights, and compare the global MPJPE against the calibrated result; if the rotation penalty no longer keeps the joint optimization near the calibrated pose, the conditioning claim fails. A simpler version: remove the background-feature rotation term and confirm that the global error jumps toward the reported 7.59 m baseline.","supporting_citations":[{"cited_title":"Akhter, T","cited_arxiv_id":null,"evidence_quote":"supplies the inverse discrete cosine transform parametrization that keeps pose and camera rotation trajectories smooth during bundle adjustment."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the 2D joint detections from image crops that the multi-view bundle adjustment reprojects."},{"cited_title":"Fasel, J","cited_arxiv_id":null,"evidence_quote":"provides the MV-Ski benchmark data and the global center-of-mass and biomechanical references used for evaluation."},{"cited_title":"Hartley and A","cited_arxiv_id":null,"evidence_quote":"establishes the homography relation between consecutive frames that turns background motion into relative camera rotations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"detects a tight bounding box around the athlete, whose interior is excluded so that only static background features are tracked."},{"cited_title":"Rhodin, J","cited_arxiv_id":null,"evidence_quote":"provides the semi-supervised monocular baseline and multi-view dataset that the paper compares against."},{"cited_title":"Rublee, V","cited_arxiv_id":null,"evidence_quote":"finds and describes the background keypoints used to estimate each camera's inter-frame homography."},{"cited_title":"Takahashi, D","cited_arxiv_id":null,"evidence_quote":"is the closest prior method that jointly optimizes human pose and uncalibrated camera rotations, which this paper's background conditioning stabilizes."}],"review_version":1}