{"id":"e4cb08f4-a05f-4718-a0f1-981081c47e0d","arxiv_id":"2509.04600","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A unified camera-and-human motion model with analytical heading decomposition and soft camera-trajectory integration improves global human trajectory reconstruction on RICH, EMDB, and 3DPW benchmarks.","lead":"This paper introduces WATCH, a system that reconstructs a person's path and body pose in world coordinates from ordinary monocular video by also tracking the camera's motion. It reports the best scores among human-motion-focused methods on standard benchmarks, which matters for VR, robotics, and automated video analysis.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (3) heading integration inherits DPVO rotation drift; Table 3's DPVO-vs-GT-gyro comparison is only two operating points and does not bound the failure regime.","rationale":"Reader's weakest assumption is the same one I would stress: Eq. (3) is an integrator and DPVO rotation is not guaranteed bias-free; the two operating points in Table 3 cannot rule out a problematic drift regime. I also note the internal evidence: Supplementary A.2 concedes numerical instability near vertical. This makes the central 'joint camera-human modeling' result conditional on a sensitivity analysis. Separately, the abstract's unqualified 'state-of-the-art' wording is contradicted by the paper's own Table 1 on EMDB-2, where TRAM and PromptHMR-vid have lower W-MPJPE100 than WATCH; the defensible claim is 'leading among human-motion-centric methods.' The paper is not fatally flawed: the decomposition is mathematically coherent, the ablation direction is consistent, and the GVHMR gains, though modest and without error bars, are plausible. The conditional verdict is appropriate; I would not accept or reject without the sensitivity check and a qualified claim.","tokens_in":16085,"tokens_out":11352,"duration_ms":108268,"concrete_test":"Run WATCH on EMDB-2 with DPVO camera increments corrupted by a constant body-frame bias: replace Delta_R_t with exp(omega_b * dt) * Delta_R_t for |omega_b| = 0.1, 0.5, and 1.0 deg/s around each camera axis, while keeping all predicted R_rp and velocities fixed, and report W-MPJPE100 and RTE as a function of bias and segment length. Separately, synthesize a camera trajectory whose forward vector pitches through vertical and measure the discontinuity in R_yaw produced by the Supplementary A.2 fallback; if the drift exceeds the GVHMR margin at realistic DPVO bias levels, or the vertical-crossing jump is more than a few degrees, the heading integrator is the weak point in the central claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central improvement over GVHMR is the recursive heading integrator in Eqs. (2)-(3): the world yaw is obtained by multiplying body-frame angular increments Delta_R_yaw,t = R_rp,t * Delta_R_cam,t * (R_rp,t+1)^T, where Delta_R_cam,t comes from DPVO (or a gyro). This is an open-loop integrator: any per-frame bias in Delta_R_yaw grows linearly in t and enters the human world orientation through Eq. (4), so the load-bearing assumption is that DPVO rotation estimates are unbiased and that the network-estimated R_rp is accurate enough for the conjugation to isolate a pure vertical-axis rotation. The paper's only evidence is Table 3, which compares two operating points (w/ DPVO and w/ GT gyro) and shows small aggregate gaps (about 1-3 mm W-A-MPJPE100 on EMDB-2). That does not bound the failure regime: DPVO errors are temporally correlated and not zero-mean, and an error in R_rp around a non-gravity axis makes the extracted 'heading' increment non-vertical, so the recursive product accumulates out-of-plane drift. Supplementary A.2 itself flags the near-vertical camera case as ill-conditioned and papers over it with a fixed fallback (Eq. 9), which can produce a discontinuous heading jump. The paper never measures how yaw error grows with sequence length or with controlled degradation of DPVO rotation, so the core assumption behind the SOTA claim is not demonstrated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces WATCH, a unified framework for world-grounded human pose and trajectory reconstruction from monocular video. The core technical contributions are an analytic heading-angle decomposition that extracts the camera's world-yaw component from body-frame angular velocity using only predicted roll and pitch, and a camera trajectory integration module that fuses camera velocity features and supervises camera and human trajectory consistency. Experiments on RICH, 3DPW, and EMDB-2 report improvements over human-motion-centric baselines such as GVHMR and WHAM, with particular gains in global trajectory metrics and smoothness. The central claim is state-of-the-art end-to-end trajectory reconstruction.","tokens_in":16389,"tokens_out":4142,"duration_ms":38159,"significance":"If the reported results hold, the work provides a simple, interpretable alternative to the view-gravity operator of GVHMR and a soft way to exploit camera translation cues without hard-decoding camera trajectories. The derivation in Section 3.2 is mathematically sound: Eq. (2) correctly isolates the heading increment by conjugation of the body-frame angular velocity, and the rotation-invariance property of Eq. (4) is valid. The teacher-forcing trajectory losses are a sensible way to decouple orientation and velocity supervision, and the commitment to release code supports reproducibility. The main open risk is that the central SOTA claim is broader than what Table 1 supports, and the recursive integrator's dependence on DPVO rotation accuracy is not stress-tested; these are fixable with additional analysis and experiments.","major_comments":[{"comment":"The abstract's unqualified claim of \"state-of-the-art performance in end-to-end trajectory reconstruction\" is not supported by Table 1: on EMDB-2, TRAM and PromptHMR-vid report W-MPJPE100 of 222.4 and 216.5 mm and RTE of 1.4 and 1.3, whereas WATCH reports 269.3 mm and 1.7. Section 4.3 narrows the claim to \"leading performance among human motion-centric methods,\" which is accurate, but the abstract and contributions section do not include that qualification. The claims should be aligned, or the comparisons with camera-trajectory-centric methods should be justified explicitly rather than deferred to qualitative physical-plausibility arguments.","section":"Abstract and Section 4.3"},{"comment":"The recursive heading integration in Eq. (3) is an open-loop integrator: any per-frame bias or temporally correlated error in the DPVO-derived body-frame angular velocity Delta_R_cam,t accumulates in the heading and propagates into the human world orientation through Eq. (4). The evidence offered in Table 3 compares only two operating points (DPVO and GT gyro) and shows small aggregate gaps, which does not bound the failure regime. The paper should quantify sensitivity to DPVO rotation errors, for example by injecting controlled rotation noise or drift, reporting heading error as a function of sequence length, and measuring the effect of the network-estimated roll-pitch error on the extracted heading. Without this, the load-bearing robustness claim in Section 4.5 is not demonstrated.","section":"Section 3.2, Eqs. (2)-(3), and Table 3"},{"comment":"The near-vertical fallback in Eq. (9) assigns a fixed heading direction when the horizontal projection of the camera forward vector is below threshold epsilon; this can produce a discontinuous jump in the heading, and therefore in the human world orientation via Eq. (4). The paper does not report how often this degenerate regime occurs in RICH or EMDB-2, nor its effect on the reported metrics. Since the method is intended for in-the-wild moving cameras, this acknowledged ill-conditioned case should be quantified or explicitly discussed as a limitation.","section":"Supplementary A.2, Eq. (9)"}],"minor_comments":[{"comment":"SLAHMR appears in both the \"Camera-trajectory-centric\" and \"Human-motion-centric\" groups with identical numbers; this is confusing and should be clarified or the method should be placed in only one category.","section":"Table 1"},{"comment":"The caption contains a typo: \"trajctory\" should be \"trajectory\".","section":"Figure 6 caption"},{"comment":"In the arXiv rendering of Table 3, the labels and method names are concatenated (e.g., \"w/ DPVOWHAM\"); the table should be reformatted so that rows and columns are clearly separated.","section":"Table 3"},{"comment":"The \"camera trajectory integration mechanism\" as described is an additive feature embedding plus an auxiliary velocity integration loss; the distinction from simple feature fusion and the claim of being \"inspired by world models\" would benefit from a more concrete explanation of why this particular integration is preferable to hard-decoding approaches.","section":"Section 3.3, Eqs. (5)-(6)"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe one thing you should know: WATCH is a competent, incremental advance on GVHMR, with a clean analytical heading decomposition and consistent wins on world-space metrics, but the paper's SOTA claim only holds inside the human-motion-centric family, and the heading integrator's dependence on DPVO rotation accuracy is not adequately stress-tested.\n\nWhat's actually new: the heading decomposition in Sec. 3.2. Writing the camera rotation as R_yaw R_rp and recovering the yaw increment by conjugating body-frame angular velocity with the roll-pitch pair is the right way to isolate heading; the math is correct and the rotational invariance property is nice. The camera trajectory integration (soft feature fusion plus auxiliary camera-velocity prediction, Eqs. 5-6) is a sensible alternative to hard-decoding. The ablations in Table 3 give real evidence that each piece helps: even without trajectory modeling WATCH slightly beats GVHMR, and adding it improves EMDB-2 W-A-MPJPE100 from 110.4 to 107.6 mm and jitter from 15.4 to 14.8.\n\nThe soft spots are mostly about scope and evidence. First, the abstract's \"state-of-the-art in end-to-end trajectory reconstruction\" is too strong; the authors themselves qualify in Sec. 4.3 to \"leading among human motion-centric methods,\" and Table 1 shows TRAM and PromptHMR-vid beating WATCH on EMDB-2 W-MPJPE100 (222.4 and 216.5 vs 269.3). Second, Eq. (3) is an open-loop integrator: any per-frame bias in DPVO's angular velocity grows linearly in time. Table 3 compares only two operating points (DPVO vs GT gyro) and shows small gaps, but that does not bound the failure regime; the supplementary's own degenerate-case fallback (Eq. 9) can produce a discontinuous heading jump. The paper never measures yaw error growth with sequence length or controlled DPVO noise. Third, no code or weights yet, and no error bars, so the 2-7 mm improvements over GVHMR should be read as likely real but not yet independently reproducible.\n\nNovelty-wise, the heading decomposition is a known fixed-axis Euler identity, and the paper acknowledges GVHMR's view-gravity operator essentially does this. So this is engineering plus a clean formulation, not a new theoretical result.\n\nVerdict: it deserves a serious referee. It is a solid, honest piece of work that should be accepted after the claims are qualified, code is released, and a robustness study of the heading integrator is added. I'd bring it to a reading group working on human motion or monocular trajectory estimation; I'd cite it if I were working in this exact niche.","headline":"Solid incremental step in global human motion reconstruction: clean heading decomposition and consistent gains over GVHMR, but the SOTA claim is only within a method family and the DPVO dependence is under-tested.","tokens_in":16936,"tokens_out":3044,"would_cite":true,"duration_ms":26059,"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":"WATCH reconstructs world-grounded human and camera trajectories from monocular video by analytically decomposing camera heading and softly integrating camera motion, reporting the top world-space accuracy among human-motion-centric…","keywords":["global human motion reconstruction","world-space human trajectory","camera-human motion coupling","heading angle decomposition","monocular video","SLAM camera trajectory","SMPL-X pose estimation"],"falsifier":"Run WATCH on a sequence with ground-truth camera and human trajectories in which the camera points nearly straight down for part of the clip; the paper's supplementary reports that this near-vertical pose makes the horizontal projection used in heading decomposition ill-conditioned. If world-space joint error spikes precisely in that segment, the analytical decomposition's stability assumption is violated. A second check: add controlled scale or drift errors to the camera angular velocity and measure W-MPJPE100 as a function of corruption; an early steep rise would indict the recursive heading integration as the bottleneck.","tokens_in":15878,"feed_emoji":"🚶","tokens_out":12152,"duration_ms":96688,"temperature":0.7,"pith_summary":"Global human motion reconstruction from ordinary monocular video fails when the camera itself moves, because the person's visible motion mixes the camera's motion with their own. This paper argues that the fix is to model both motions jointly: it introduces WATCH, a single network that reconstructs the camera and the human in a shared world frame. The method's central move is to reduce the hardest orientation problem—the camera's heading around gravity—to an analytical formula that needs only the camera's roll and pitch, obtained by recursive integration of the camera's body-frame angular velocity. It then feeds camera velocity into the model as soft context rather than as a hard positional constraint, avoiding the pose discontinuities typical of direct SLAM-to-body mapping. On in-the-wild benchmarks, WATCH reports the best world-space joint errors, smoothness, and foot-contact quality among human-motion-centric methods, indicating that explicit camera-human coupling is a viable path to stable global trajectories.","feed_headline":"Modeling camera and human together cuts world-space trajectory error","feed_subtitle":"Analytic heading decomposition plus soft camera cues beat prior human-only methods on RICH and EMDB.","key_machinery":"The load-bearing object is the analytical heading decomposition. It uses a fixed-axis Euler decomposition in a gravity-aligned world frame, $R^{\\mathrm{cam}} = R^{\\mathrm{cam}}_{\\mathrm{yaw}} R^{\\mathrm{cam}}_{\\mathrm{rp}}$, and isolates heading changes by conjugating the body-frame angular velocity with the roll-pitch component: $\\Delta R^{\\mathrm{cam}}_{\\mathrm{yaw},t} = R^{\\mathrm{cam}}_{\\mathrm{rp},t} \\Delta R^{\\mathrm{cam}}_t (R^{\\mathrm{cam}}_{\\mathrm{rp},t+1})^T$. This matters because it converts heading estimation into a recursion: predict only roll-pitch, then multiply the heading increments and obtain both camera and human world orientations. The second mechanism is the camera trajectory integration module: camera local velocity is fused into the embedding space through MLPs, and an auxiliary decoder predicts camera local velocity and integrates it into a trajectory under teacher-forcing consistency losses, so translation cues shape the human trajectory without imposing direct depth constraints.","core_discovery":"On the paper's own terms, the central discovery is that world-space human orientation can be recovered without estimating heading directly. WATCH decomposes the camera rotation into a heading (yaw) part and a roll-pitch part, $R^{\\mathrm{cam}}_t = R^{\\mathrm{cam}}_{\\mathrm{yaw},t} R^{\\mathrm{cam}}_{\\mathrm{rp},t}$. Because the body-frame angular velocity $\\Delta R^{\\mathrm{cam}}_t$ measures rotation changes in the camera's local coordinates, the heading increment can be isolated analytically: $\\Delta R^{\\mathrm{cam}}_{\\mathrm{yaw},t} = R^{\\mathrm{cam}}_{\\mathrm{rp},t} \\Delta R^{\\mathrm{cam}}_t (R^{\\mathrm{cam}}_{\\mathrm{rp},t+1})^T$. So the network only needs to predict the more intuitive roll-pitch component, and the heading follows by integration; combining it with the camera-space human orientation gives the world-space orientation $R^{\\mathrm{h,w}}_t = R^{\\mathrm{cam}}_{\\mathrm{yaw},t} R^{\\mathrm{cam}}_{\\mathrm{rp},t} R^{\\mathrm{h,c}}_t$. For translation, the paper deliberately avoids hard-decoding camera trajectories. Instead, camera local velocity is encoded as a soft contextual feature, the decoder predicts both human and camera local velocities, and trajectories are produced by integration with teacher-forcing consistency losses. The combined effect, according to the paper's experiments, is lower world-aligned joint error and less jitter and foot sliding than previous human-motion-centric approaches, while retaining camera-space accuracy.","pith_inferences":["Going beyond the paper: because heading is constructed analytically, one could calibrate the roll-pitch network with cheap IMU gravity measurements instead of full camera-pose ground truth, easing deployment on unseen handheld rigs.","Going beyond the paper: feeding the model deliberately corrupted camera angular velocities with controlled scale and drift errors and measuring world-space joint error as a function of corruption would isolate how much of the gain depends on odometry accuracy, a sensitivity analysis the paper does not report.","Going beyond the paper: the soft camera-trajectory module is architecture-agnostic enough that it could be grafted onto other human-centric baselines; if it transfers, camera-translation integration would become a plug-in competence rather than a bespoke design.","Going beyond the paper: the near-vertical camera pose case flagged in the supplementary suggests that augmenting training with downward- and upward-looking footage, or reparameterizing heading for that regime, would be a natural stress test."],"forward_implications":["If the central claim holds, any human-motion-centric pipeline can adopt the heading decomposition to obtain world-space orientation from only roll-pitch and angular velocity, making camera-orientation cues cheaper and more interpretable than geometric projection operators.","Soft camera-trajectory integration should improve global trajectory accuracy and smoothness without the pose discontinuities that direct SLAM-to-human mapping produces.","The benefit is largest when the camera moves; on the static-camera RICH setting, the paper reports camera-space results slightly below the strongest baseline, so the gain is tied to dynamic cameras.","The framework retains its advantage when camera motion comes from odometry as well as from an ideal gyroscope, so the method is intended to transfer to realistic SLAM-derived inputs.","Joint camera and human velocity decoding gives a route to end-to-end training that avoids cascading errors from separately estimated camera trajectories over long sequences."],"supporting_citations":[{"why":"Defines the human-motion-centric baseline and the view-gravity operator that WATCH replaces with analytical heading decomposition.","marker":"GVHMR (Shen et al. 2024)"},{"why":"Establishes the human-velocity-centric paradigm and the evaluation framework on which WATCH builds.","marker":"WHAM (Shin et al. 2023)"},{"why":"Provides the body-frame camera angular velocity and camera extrinsics used in the realistic odometry-setting experiments.","marker":"DPVO (Teed et al. 2023)"},{"why":"Supplies the moving-camera benchmark with ground-truth global trajectories used for the main world-space comparison.","marker":"EMDB (Kaufmann et al. 2023)"},{"why":"Supplies the static-camera benchmark with ground-truth global human motion for world-space evaluation.","marker":"RICH (Huang et al. 2022)"},{"why":"Frozen ViT encoder whose image features condition the camera-space and trajectory predictors.","marker":"HMR2.0 (Goel et al. 2023)"},{"why":"A camera-trajectory-centric hard-decoding baseline that WATCH contrasts with on physical plausibility.","marker":"TRAM (Wang et al. 2024b)"},{"why":"A camera-trajectory-centric decoupling baseline and source of the W-MPJPE evaluation protocol.","marker":"SLAHMR (Ye et al. 2023a)"}],"fun_headline_variants":["Heading solved analytically; soft camera cues cut world error","Skip heading prediction: decompose camera yaw from roll-pitch","WATCH: joint camera-human modeling wins global pose benchmarks","Soft camera cues beat hard decoding in world-space human tracking","Analytic yaw recovery: no direct heading estimate needed"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the camera's turning-rate estimates from odometry or a gyroscope are accurate enough that repeatedly adding up the extracted heading does not drift, and that the network's estimate of how the camera tilts relative to gravity is correct; if either fails, the person's world-space orientation and trajectory degrade.","fun_headline_variants_meta":{"raw":{"variants":["Heading solved analytically; soft camera cues cut world error","Skip heading prediction: decompose camera yaw from roll-pitch","WATCH: joint camera-human modeling wins global pose benchmarks","Soft camera cues beat hard decoding in world-space human tracking","Analytic yaw recovery: no direct heading estimate needed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000204,"raw_usage":{"total_tokens":1456,"prompt_tokens":1080,"completion_tokens":376,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":696,"completion_tokens_details":{"reasoning_tokens":295}},"tokens_in":696,"tokens_out":376,"duration_ms":3812,"temperature":1.0,"reasoning_tokens":295,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:29:12.740692+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run WATCH on a sequence with ground-truth camera and human trajectories in which the camera points nearly straight down for part of the clip; the paper's supplementary reports that this near-vertical pose makes the horizontal projection used in heading decomposition ill-conditioned. If world-space joint error spikes precisely in that segment, the analytical decomposition's stability assumption is violated. A second check: add controlled scale or drift errors to the camera angular velocity and measure W-MPJPE100 as a function of corruption; an early steep rise would indict the recursive heading integration as the bottleneck.","supporting_citations":[],"review_version":2}