ManiPose: Manifold-Constrained Multi-Hypothesis 3D Human Pose Estimation

Authors: Cédric ROMMEL, Victor Letzelter, Nermin Samet, Renaud Marlet, Matthieu Cord, Patrick Perez, Eduardo Valle

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We showcase the performance of Mani Pose on real-world datasets, where it outperforms state-of-the-art models in pose consistency by a large margin while being very competitive on the MPJPE metric.Extensive empirical results, including comparison to strong baselines, evaluation on two challenging datasets (Human 3.6M and MPI-INF-3DHP), and ablations.
Researcher Affiliation Collaboration 1Valeo.ai, Paris, France 2Sorbonne Université, Paris, France 3LTCI, Télécom Paris, Institut Polytechnique de Paris, France 4Recod.ai Lab, School of Electrical and Computing Engineering, University of Campinas, Brazil 5LIGM, Ecole des Ponts, Univ Gustave Eiffel, CNRS, Marne-la-Vallee, France
Pseudocode Yes Algorithm 1 6D rotation representation conversion [54]... Algorithm 2 Forward Kinematics [34, 26]
Open Source Code Yes The Py Torch [37] implementation of Mani Pose and code used for all our experiments can be found at https://github.com/cedricrommel/manipose.
Open Datasets Yes Human 3.6M [15] contains 3.6 million images of 7 actors performing 15 different indoor actions. It is the most widely used dataset for 3D-HPE.Dataset licences. Human 3.6M is a dataset released under a research-only custom license, and is available upon request at this URL: http://vision.imar.ro/human3.6m/description.php. MPI-INF-3DHP is released under non-commercial custom license and can be found at: https://vcai.mpi-inf.mpg.de/3dhp-dataset/.
Dataset Splits Yes We created a dataset of input-output pairs {(xi, (xi, yi))}N i=1, divided into 1 000 training examples, 1 000 validation examples and 1 000 test examples.
Hardware Specification Yes Trainings were carried out on a single NVIDIA RTX 2000 GPU with around 11GB of memory.
Software Dependencies No The paper mentions 'Py Torch' but does not specify a version number or other software dependencies with version numbers.
Experiment Setup Yes We trained our model for a maximum of 200 epochs with the Adam optimizer [17], using default hyperparameters, a weight decay of 10 6 and an initial learning rate of 4 10 5. The latter was reduced with a plateau scheduler of factor 0.5, patience of 11 epochs and threshold of 0.1 mm. Batches contained 3 sequences of T = 243 frames each for the Human 3.6M training, and 30 sequences of T = 43 frames for MPI-INF-3DHP.