Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
3D Human Pose Estimation with Muscles
Authors: Kevin Zhu, AliAsghar MohammadiNasrabadi, Alexander Wong, John McPhee
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We show that Muscle Pose is competitive with existing 3D pose estimators in positional accuracy, while also able to infer plausible human kinetics and muscle signals consistent with values from biomechanics studies, without requiring an external physics engine. ... We demonstrate improvements in the inferred kinetics on actions including walking from the H36M dataset [27] and baseball pitching and golf swings from Penn Action [84]. ... 4 Experiments |
| Researcher Affiliation | Academia | Kevin Zhu Ali Asghar Mohammadi Nasrabadi Alexander Wong John Mc Phee University of Waterloo EMAIL |
| Pseudocode | No | The paper describes methods using mathematical equations and textual descriptions, but it does not contain any clearly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | No | Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] Justification: while our code is not yet ready for open source, we included in the supplementary material, all necessary details required to reproduce our model and results, including implementation details of all novel components, as well as all the publicly available code repositories we borrowed from to implement and train our model, and to produce and evaluate our results. |
| Open Datasets | Yes | We demonstrate improvements in the inferred kinetics on actions including walking from the H36M dataset [27] and baseball pitching and golf swings from Penn Action [84]. ... For training, we used the AMASS dataset [46], with feet-ground contact labels from [83]. ... For evaluation, we assessed positional accuracy on the inference results of the H36M test set [27] and object-occlusion subset of 3DPW (3DPWoc) [71]. |
| Dataset Splits | Yes | For training, we used the AMASS dataset [46], with feet-ground contact labels from [83]. As such, we trained and evaluated on sequences with feet-ground contact only (denoted ), which is also the case for many PHPE experiments [37, 20, 82]. ... For evaluation, we assessed positional accuracy on the inference results of the H36M test set [27] and object-occlusion subset of 3DPW (3DPWoc) [71]. |
| Hardware Specification | Yes | The entire process can be trained in about 12 hours on a single Titan Xp GPU. |
| Software Dependencies | No | We used the Adam W optimizer [44] with a weight decay of 10 4 and an initial learning rate of 10 4 that decreases by 20% every 5 epochs. ... Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? Answer: [Yes] Justification: all models implemented can be found in Sec. 3, with details, such as model coefficients or hyperparameters, listed in the supplementary material. |
| Experiment Setup | Yes | We trained Muscle Pose end-to-end with a sequence input length of 16 frames, using total loss Ltotal = Lkin + Ldyn for 25 epochs, using the Adam W optimizer [44] with a weight decay of 10 4 and an initial learning rate of 10 4 that decreases by 20% every 5 epochs. Following common curriculum learning [2] practices, we split the training into two phases for the first 20 epochs, we trained using the ground truth as input, followed by 5 epochs using the model s predictions as inputs. |