{"id":"29cff4fb-aa9b-4efa-92e9-ce2b187bc012","arxiv_id":"2501.18478","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"SimpleDepthPose estimates 3D human poses from multiple RGBD cameras by extracting depth at visible 2D joints and merging the resulting proposals, without any neural 3D training.","lead":"This paper presents SimpleDepthPose, a fast training-free algorithm for multi-view, multi-person 3D pose estimation that combines 2D keypoints with depth images instead of learned triangulation. The method detects more people and keypoints than existing approaches in heavily occluded scenes, while running faster than most.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'best generalization' claim rests on baseline numbers taken from the authors' own prior paper [1] without independent reproduction; the reported 20-point PCP lead over VoxelKeypointFusion is not yet verified.","rationale":"The reader's weakest_assumption focuses on the requirement that the 2D detector suppress occluded keypoints and on the static per-joint offsets. These are real limitations and the paper itself acknowledges them (Section 3, step 1; Section 5 ablation shows visibility finetuning strongly reduces invalid predictions). However, they are internally handled for the main 13-keypoint evaluation: the authors use a finetuned HigherHRNet, and the ablation supports the necessity. The more load-bearing issue for the central comparative claim is that the baseline numbers are taken from the same group's earlier paper without independent verification. If those numbers are not reproduced under identical conditions, the claim of being 'best' is not established. This is not an accusation of dishonesty; it is a matter of evidence quality. The paper does make its source code public, so independent reproduction is feasible, which is why the verdict should remain CONDITIONAL rather than REJECT: the claim may well survive the check, but it currently depends on unverified comparisons. The concrete test above would settle whether the reported performance advantage is genuine or an artifact of evaluation protocol. I thus agree with the reader's overall CONDITIONAL verdict but not with the specific weakest_assumption emphasized in their stress-test.","tokens_in":8612,"tokens_out":11969,"duration_ms":129146,"concrete_test":"Independently reproduce the most relevant RGBD baselines on MVOR: run VoxelKeypointFusion (code from [1]) and OpenPTrack using the official evaluation script, and feed both methods the same HigherHRNet 2D detections (with visibility finetuning) that SimpleDepthPose uses. Measure PCP, Recall@500, Invalid, and F1. If SimpleDepthPose's PCP lead over VoxelKeypointFusion drops below 10 points, or if VoxelKeypointFusion's with-depth PCP rises above 60, the 'best generalization' claim in Table 1 is not robust. If PointVoxel's code becomes available, also evaluate it on MVOR and check whether it surpasses SimpleDepthPose.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is comparative: SimpleDepthPose 'shows the best generalization results among other methods' (Conclusion) and outperforms all listed baselines on MVOR by a wide margin (Table 1: PCP 74.0 vs. 54.0 for VoxelKeypointFusion with depth). However, the table caption states 'All other results without extra citations are taken from [1]' — that is, from the authors' own earlier VoxelKeypointFusion paper. The baselines are not re-run in this work, no error bars or variance are reported, and no evidence is provided that the baselines used the same 2D pose detections, the same evaluation script, or the same hyperparameters as SimpleDepthPose. This matters because SimpleDepthPose and VoxelKeypointFusion share the same 2D detection stage; if the baseline numbers were produced under a slightly different protocol (e.g., different keypoint filtering, stricter person-visibility criterion, or different depth alignment), the large reported gap could shrink substantially. Additionally, PointVoxel [16], a recent RGBD method aimed at cross-dataset generalization, is mentioned but not evaluated because its code is unavailable. Since the conclusion is entirely comparative, the absence of an independent, third-party reproduction of the key baseline numbers makes the headline claim unverified rather than established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents SimpleDepthPose, a training-free algorithm for multi-view, multi-person 3D pose estimation from RGBD images. The method uses an off-the-shelf 2D pose estimator, extracts depth values from aligned depth images for each visible joint, groups per-view 3D proposals into persons across time, filters outliers with limb-length heuristics, and averages the remaining proposals. The authors evaluate the method on the MVOR and Panoptic datasets, reporting strong detection rates especially under occlusions, competitive accuracy, and high frame rates. They also provide ablations on camera count, visibility finetuning, joint offsets, and alternative point-cloud fusion strategies. The central claim is that SimpleDepthPose generalizes better than prior methods without any additional training, while being fast and reliable.","tokens_in":8932,"tokens_out":3892,"duration_ms":40111,"significance":"If the reported results are independently reproduced, this is a practically valuable contribution: it is a simple, deterministic, training-free baseline that requires no learned 3D fusion and runs at real-time speed. The public release of code is a concrete strength, as is the fact that the method fits no parameters to the evaluation data, reducing circularity concerns. The paper's ablations are informative, particularly the camera-count study and the analysis of visibility finetuning. However, the comparative claim of 'best generalization' rests on baseline numbers taken mostly from the authors' own prior paper, with no error bars or protocol-equivalence evidence; the method also depends on load-bearing assumptions (visible-joint-only 2D detection, static per-joint offsets) that are not quantitatively stress-tested. The contribution is therefore plausible but not yet fully established.","major_comments":[{"comment":"The baseline numbers without extra citations are taken from the authors' own prior work [1], as stated in the table captions. Since the paper's headline conclusion is explicitly comparative ('shows the best generalization results among other methods'), the absence of independent re-runs or a detailed equivalence argument is load-bearing. The reported 20-point PCP lead over VoxelKeypointFusion on MVOR could shrink if the baselines were evaluated under slightly different keypoint filtering, person-visibility criteria, or depth alignment. Please provide either a same-protocol re-run of all baselines, or a precise specification of what was shared (2D detections, evaluation script, thresholds) and an analysis of how protocol choices affect the comparison. Reporting variances or error bars would additionally allow an assessment of significance.","section":"Table 1 and Table 2"},{"comment":"The method explicitly requires that the 2D pose estimator predict only directly visible keypoints, because occluded joints produce meaningless depth values. This is a load-bearing assumption, yet the paper offers no quantitative evidence of how often the visibility-finetuned HigherHRNet model emits occluded joints on MVOR or Panoptic, nor how the outlier filter behaves when this assumption fails. Please report the precision of the visible-joint prediction (e.g., the fraction of emitted keypoints that are actually visible in the depth image) and include an ablation with a standard COCO detector without visibility finetuning, so that the sensitivity of the downstream method to this requirement is quantified.","section":"Section 3, Step (1)"},{"comment":"The static per-joint depth offsets (e.g., 3 cm for shoulders/knees, 1 cm for wrists) are estimated from 'normal human proportions' and can be adapted, but no sensitivity analysis is provided. Since the method is positioned as reliable in varied settings, including thick clothing or unusual body proportions, the impact of these offsets on MPJPE and PCP should be quantified. A simple ablation varying the offsets (or a discussion of the expected error range) would indicate how much the reported accuracy depends on these hand-set values.","section":"Section 3, Step (2)"},{"comment":"No error bars, confidence intervals, or repeated-run variances are reported for any metric. Because the central claims are comparative (higher recall, higher F1, faster than other methods), single-number entries without variance information make it impossible to judge whether differences are meaningful. Please add variance information over evaluation videos or subjects, or at least state the number of frames used for each metric and dataset.","section":"Tables 1–3"}],"minor_comments":[{"comment":"The cross-shape depth extraction is described textually and in Figure 2, but the dimensions of the cross (pixel sizes, arm lengths) are never specified; please give concrete numbers.","section":"Section 3, Step (2)"},{"comment":"'operation room' should be 'operating room' for consistency with standard terminology.","section":"Section 4"},{"comment":"The abstract claims the method is 'adaptable to different keypoints', but Section 6 shows that the visibility finetuning cannot be applied to face and finger keypoints, and whole-body performance is limited. Please temper the claim or report quantitative whole-body results.","section":"Abstract and Section 6"},{"comment":"The two VoxelKeypointFusion rows (one without depth, one with depth) are easy to confuse; consider labeling the rows explicitly as 'RGB' and 'RGBD'.","section":"Tables 1 and 2"},{"comment":"The name 'HigherHrNet' is written with inconsistent capitalization; use 'HigherHRNet' consistently.","section":"Throughout"},{"comment":"The source-code URL is a bare repository link; please provide a specific commit or version for reproducibility.","section":"Section 1"}],"recommendation":"major_revision","confidential_remarks":"The heavy reliance on the authors' own prior paper [1] for almost all baseline numbers is a concern both for verification and for the novelty framing; an independent evaluation or a clearly scoped comparison protocol is needed. The paper is within the journal's scope, and the method itself is a valuable deterministic baseline. If the authors can provide the requested sensitivity analyses and a more rigorous comparative setup, the contribution would be solid."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"SimpleDepthPose is a genuinely simple, training-free pipeline for multi-view multi-person 3D pose estimation from RGBD images, and the reported improvement in detection rate on heavily occluded scenes looks real, though the exact size of the advantage is not yet verified because the headline baseline numbers come from the authors' own prior VoxelKeypointFusion paper.\n\nWhat is actually new: the specific combination of a visibility-only 2D detector, per-joint median depth extraction using a cross-shaped window, static anthropometric depth offsets, neighbor-based outlier rejection, and top-k averaging across views. Individual components are not novel—OpenPTrack used the depth-extraction-plus-fusion idea—but the packaging is clean, the ablations are informative, and it works without training. The paper is honest about the key requirement: the 2D detector must output only visible joints, otherwise depth values are wrong. They finetune HigherHRNet for that, and show the effect in ablations. Code is public and the FPS numbers are measured, which is more than many papers in this area offer.\n\nThe soft spot is the comparative evaluation. Table 1 states that all other results without citations are taken from [1], the authors' earlier paper. None of the baselines are re-run here, there are no error bars, and we don't know if the same 2D detections, evaluation script, and hyperparameters were used. So the 20-point PCP lead over VoxelKeypointFusion is plausible but unverified. The stress-test note gets this right. This matters because the conclusion is entirely comparative. A fair side-by-side could shrink the gap substantially. Also, the method inherits the limitations of the 2D detector and the static offsets; thick clothing or unusual body proportions will bias the joints. These are real caveats, not fatal ones.\n\nThe method itself is parameter-free and not fit to the test data, so circularity is not a concern. The qualitative conclusion—depth helps detect more people in occlusion, at some cost in localization accuracy—is supported by the experiments.\n\nThis paper is for practitioners and researchers working on multi-camera RGBD pose estimation, especially in robotics or operating rooms. It deserves a serious referee. The right revision would re-run or provide third-party baselines, add error bars, and discuss the detector dependency. I would cite it if I worked in this area.\n\nRecommendation: engage with the work; the evaluation needs strengthening, but the method and the ablations are worth referee time.","headline":"Simple, training-free RGBD pose estimator with plausible gains on occluded scenes; headline comparison rests on inherited baseline numbers, but the method itself is honest and reproducible.","tokens_in":9425,"tokens_out":2518,"would_cite":true,"duration_ms":23054,"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":"SimpleDepthPose estimates 3D human poses by reading depth values at visible 2D joints and averaging them, with no learned triangulation and no additional training.","keywords":["multi-view pose estimation","RGBD images","depth-based 3D pose","multi-person pose estimation","no-training generalization","outlier filtering","PCP metric","depth extraction"],"falsifier":"Run SimpleDepthPose on a multi-view RGBD sequence using a 2D pose estimator that does not suppress occluded keypoints, such as the default COCO-trained HigherHRNet with its refinement step kept, and compare the F1 and MPJPE to the same pipeline with visibility-finetuned keypoints; a sharp rise in invalid predictions and joint errors would confirm that the visible-only assumption carries the result. Alternatively, test on subjects wearing thick or loose clothing and measure the systematic offset in joint positions.","tokens_in":8401,"feed_emoji":"🧍","tokens_out":5563,"duration_ms":51931,"temperature":0.7,"pith_summary":"This paper introduces SimpleDepthPose, a method for estimating the 3D poses of multiple people from several RGBD cameras. It claims that reliable multi-view pose estimation can be achieved without learned triangulation or any additional training: take off-the-shelf 2D joint detections from each color image, read the depth at each visible joint, and fuse the resulting 3D proposals by grouping, outlier filtering, and averaging. On the MVOR operating-room dataset it reports a PCP (percentage of correct parts) of 74.0, recall of 96.6, and F1 of 85.4, higher than all compared baselines, while running at 37.2 frames per second. The value of the claim is practical: safety-critical applications such as human-robot collaboration need fast, reliable pose estimates, and depth cameras are increasingly available.","feed_headline":"Depth extraction tops learned 3D pose baselines at 37 FPS","feed_subtitle":"Depth values replace learned triangulation, so occluded people are detected even from a single view.","key_machinery":"The central mechanism is a pipeline of five simple steps: 2D pose prediction per view, depth extraction, person grouping, outlier filtering, and fusion by averaging. The depth extraction uses a cross-shaped set of pixels around each detected joint and takes the median value, which suppresses outliers at object edges; the per-joint offsets account for the fact that a depth camera measures the surface while the target is the joint center. The outlier filter checks each joint proposal against the averaged center of its neighboring joints and discards proposals with implausibly long limbs. The grouping step assigns each 3D proposal to the closest person from the previous time step, creating new persons when needed. Together these steps let the method skip the learned voxel or triangulation networks used by other approaches.","core_discovery":"On the paper's own terms, the central discovery is that depth information can replace the learnable triangulation step entirely. For each color image the method predicts only directly visible 2D keypoints, reads the median depth from a cross-shaped window around each keypoint, adds a static per-joint offset (for example 3 cm for shoulders and knees, 1 cm for wrists) to reach the joint center, and transforms the resulting camera-space points into world coordinates. Person proposals are grouped across views and over time, outliers whose limb lengths exceed a threshold are discarded, and the remaining proposals are averaged. The result is a detection rate and accuracy that the paper reports as above all baselines on the MVOR dataset and comparable to learned methods on the Panoptic dataset, at 37.2 FPS on a single Nvidia-3090. In strongly occluded settings, a single view with depth is enough to detect a person, which triangulation-based methods cannot do.","pith_inferences":["An implicit consequence is that the method's accuracy is capped by the depth sensor's noise and calibration; future work could couple it with a lightweight learned refinement of the joint offsets rather than static proportions.","The visibility-only requirement suggests a natural test for 2D detectors: the same pipeline could serve as a probe for whether a detector suppresses occluded keypoints, since occluded outputs will show up as large limb-length outliers.","The point-cloud fusion experiments indicate the method can consume depth from non-camera sources, which would matter for robotics setups that already produce fused point clouds.","A temporal extension with a simple Kalman filter, avoided here to prevent lag, could be reintroduced selectively for slow-moving joints to improve accuracy without hurting fast motions."],"forward_implications":["If the reported results hold, depth sensors can replace learned triangulation networks, removing the need for per-scene training data in multi-view pose estimation.","In occluded scenes, a person visible in only one camera can still be detected and posed, because depth provides the third dimension from a single view.","The method runs at 37.2 FPS on an Nvidia-3090, making it suitable for real-time and safety-critical applications such as human-robot collaboration.","Because the 3D pipeline is training-free, it can be quickly adapted to other keypoint sets, as demonstrated with whole-body keypoints.","The ablation of camera count shows performance degrades gracefully with fewer cameras, so the approach can be tuned to available hardware."],"supporting_citations":[{"why":"Supplies the evaluation metrics, the baseline results in the comparison tables, and the algorithmic point of comparison for keypoint adaptability.","marker":"[1]"},{"why":"Provides the HigherHRNet 2D pose detector that is finetuned to predict only visible joints, a central requirement of the method.","marker":"[3]"},{"why":"The MVOR operating-room dataset used for the main transfer evaluation, where occlusions are heavy.","marker":"[19]"},{"why":"The Panoptic dataset used for the second evaluation, providing RGBD images and 3D pose labels.","marker":"[10]"},{"why":"Describes OpenPTrack, the closest architectural baseline that also extracts depth at 2D joints and uses a Kalman filter; the paper compares against it directly.","marker":"[2]"},{"why":"The COCO dataset on which the 2D detector is trained and finetuned to predict only visible joints.","marker":"[13]"},{"why":"Provides the RTMPose whole-body keypoint model used to extend the method to face, hand, and foot keypoints.","marker":"[9]"}],"fun_headline_variants":["Depth replaces triangulation for 3D pose at 37 FPS","Single-view depth finds occluded people in 3D pose","Fast 3D pose: depth instead of learned triangulation","No learned triangulation: depth-only 3D pose at 37 FPS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes the 2D pose estimator outputs only directly visible joints and never occluded ones, because an occluded joint has no meaningful depth value; if this fails, wrong 3D joints and reduced detection accuracy follow.","fun_headline_variants_meta":{"raw":{"variants":["Depth replaces triangulation for 3D pose at 37 FPS","Single-view depth finds occluded people in 3D pose","Fast 3D pose: depth instead of learned triangulation","No learned triangulation: depth-only 3D pose at 37 FPS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000208,"raw_usage":{"total_tokens":1338,"prompt_tokens":811,"completion_tokens":527,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":427,"completion_tokens_details":{"reasoning_tokens":450}},"tokens_in":427,"tokens_out":527,"duration_ms":5510,"temperature":1.0,"reasoning_tokens":450,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T23:19:20.340327+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SimpleDepthPose on a multi-view RGBD sequence using a 2D pose estimator that does not suppress occluded keypoints, such as the default COCO-trained HigherHRNet with its refinement step kept, and compare the F1 and MPJPE to the same pipeline with visibility-finetuned keypoints; a sharp rise in invalid predictions and joint errors would confirm that the visible-only assumption carries the result. Alternatively, test on subjects wearing thick or loose clothing and measure the systematic offset in joint positions.","supporting_citations":[{"cited_title":"Higherhrnet: Scale- aware representation learning for bottom-up human pose es- timation","cited_arxiv_id":null,"evidence_quote":"Provides the HigherHRNet 2D pose detector that is finetuned to predict only visible joints, a central requirement of the method."},{"cited_title":"Real-time marker-less multi-person 3D pose es- timation in RGB-depth camera networks","cited_arxiv_id":null,"evidence_quote":"Describes OpenPTrack, the closest architectural baseline that also extracts depth at 2D joints and uses a Kalman filter; the paper compares against it directly."}],"review_version":1}