Prosthesis-Aware 3D Human Pose Estimation: A Dataset and Benchmark for RSP Users
Yilin Wen, Kechuan Dong, Fumiya Suginaka, Ken Endo, Yusuke Sugano
cs.CV
Quick verdict
First 3D dataset of RSP prosthesis users with zero-shot benchmark; novel and careful resource, but 6 subjects and no RSP gain over model-free limit impact.
Summary
The paper introduces prosthesis-aware 3D human pose estimation: recovering both intact body joints and running-specific prosthesis (RSP) geometry from monocular video. The authors collect RSP3D, the first 3D dataset of RSP users: 6 participants (below-knee, above-knee, and bilateral amputations; ages 10s-60s) perform about 20 daily-life and exercise actions recorded by 16 GoPro cameras synchronized via QR code (within one frame at 60 FPS) and calibrated against an OptiTrack marker system, yielding 5.6M frames with Human3.6M-style 17-joint annotations plus an RSP surface reconstructed as a ruled surface through paired marker chains. Zero-shot evaluation shows model-based estimators (best: SAM3D Body at 78.35mm MPJPE) produce accurate joints but cannot represent prosthesis geometry and hallucinate amputated limbs, while model-free SpatialTrackerV2 recovers RSP points but with poor joint accuracy (~120-175mm) and heavy sensitivity to 2D detection quality. A hybrid baseline takes SAM3D Body joints and aligns model-free RSP points via pelvis-rooted scale correction, matching oracle-scale model-free RSP accuracy (CD 251-288mm) while retaining model-based joint accuracy. Main limitations: six subjects cannot support training or variance estimates; each segment uses one unspecified camera view; body-space RSP accuracy remains poor, especially for bilateral amputees where hybrid CD (342mm) is worse than pure model-free (294mm).
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- RSP3D is the first 3D dataset of running-specific prosthesis users, providing 5.6M frames with both body-joint and RSP-surface annotations from 6 participants with varied amputation conditions.
- Zero-shot evaluation confirms that model-based estimators produce accurate joints but cannot represent prosthesis geometry, while model-free methods recover arbitrary shapes but lack body kinematic priors and fail under occlusion.
- A hybrid baseline combining SAM3D Body joints with scale-aligned SpatialTrackerV2 RSP points retains model-based joint accuracy (78.35mm MPJPE) while matching oracle-scale model-free RSP-shape accuracy.
- 2D detection and segmentation quality for prosthesis users is a significant bottleneck, as replacing detections with 2D ground truth substantially improves all model-free results.
- Participants with bilateral amputations show higher estimation errors due to fewer intact joints available as anchors and the difficulty of aligning two prostheses to the body.
Strengths
First-of-kind dataset with dual 3D annotation
RSP3D provides 5.6M frames with both Human3.6M-style 17-joint body annotations and an RSP surface (ruled surface over paired marker chains, Eq. 1); Table 1 shows no prior dataset combines 3D prosthesis and 3D body annotations, and participants span below-knee, above-knee, and bilateral amputations (Table 2).
Careful and well-documented capture pipeline
Section 3.1 details QR-video time synchronization (verified within one frame at 60 FPS, following Ego-Exo4D), ChArUco-based intrinsic/extrinsic calibration tied to the OptiTrack world frame; Section 3.2 describes the action list co-designed with two experienced prosthetists after observing a first-time RSP training session, with full per-action protocols in supplementary Section B.
Metric design that isolates the actual failure mode
Transparent, fine-grained error analysis
Oracle ablations localize the bottleneck
Weaknesses
Six participants; zero-shot-only benchmark
Hybrid baseline does not improve RSP-shape metrics
Unspecified single-view selection protocol
Ground-truth dependencies inside a 'zero-shot' benchmark
Coarse RSP ground-truth representation
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the RSP3D dataset?
RSP3D is described as the first 3D motion dataset of running-specific prosthesis (RSP) users. Six amputee participants (below-knee, above-knee, and bilateral amputations) performed about 20 daily-life and exercise actions in a lab equipped with 16 GoPro cameras and a marker-based OptiTrack system, producing 5.6 million frames annotated with 3D positions of both natural body joints and the prosthesis surface.
Why can't existing 3D human pose estimators handle prosthesis users?
Model-based methods are trained on body models of non-amputee people, so they hallucinate a missing limb and cannot represent prosthesis geometry at all. Model-free 3D reconstruction methods can capture arbitrary shapes, but without body kinematic priors they estimate joints unreliably, especially under occlusion. This paper shows both failure modes empirically on prosthesis users.
How well do current methods perform on RSP3D?
In zero-shot tests, the best model-based method (SAM3D Body) reaches 78.35mm mean per-joint error but outputs no prosthesis geometry. The model-free SpatialTrackerV2 recovers prosthesis points with roughly 250-300mm Chamfer distance but much worse joints (about 120-175mm). The paper's hybrid baseline keeps the 78.35mm joint accuracy and matches model-free-level prosthesis accuracy, though body-relative prosthesis placement remains inaccurate, especially for bilateral amputees.
Is the RSP3D dataset publicly available?
According to the paper, data and code are accessible via the project page (ut-vision.github.io/RSP3D), but access is gated: users must agree to a license and terms of use and submit an access request form, after which approved applicants receive the videos and annotations.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- All figures (Figs. 1-8) are available only as captions in the extracted text; qualitative judgments rely on the authors' descriptions of the visualizations.
- The supplementary video (supp.mp4) referenced in the text cannot be reviewed from the text extraction.
- The exact number and list of evaluated action segments per participant is not stated in the extracted text, limiting assessment of evaluation scope.
- Table values were extracted as text and appear internally consistent, but OCR-related digit errors cannot be fully ruled out.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.