library_name: pytorch
tags:
- motion-capture
- smplh
license: other
license_name: dirtymocap-noncommercial-research
license_link: https://huggingface.co/aLongZJU/DirtyMoCap/blob/main/LICENSE
DirtyMoCap
Inference weights for DirtyMoCap.
License
Both v1 and v-dataset are governed by the same pretrained model license below.
The provided checkpoints may be used only for non-commercial scientific research, non-commercial education or non-commercial artistic projects, under the DirtyMoCap Pretrained Model License. This condition also applies to use of modified or fine-tuned versions.
Training used AMASS and GRAB motion data together with SMPL-H/MANO. The v-dataset version additionally used BEDLAM2 training data. Users must comply with the applicable AMASS, GRAB, BEDLAM2 (for v-dataset) and SMPL-H/MANO terms. Body models and other third-party assets must be obtained separately under their own licenses.
The original DirtyMoCap source code is licensed separately under Apache 2.0. That source-code license does not cover these checkpoints or third-party resources.
v1 checkpoints
Use both checkpoints from the same revision with the matching DirtyMoCap inference code and configuration.
| File | Component |
|---|---|
| v1/init.ckpt | Init network |
| v1/refine_solver.ckpt | Refine/Solver network |
| v1/SHA256SUMS | SHA256 checksums |
| v1/LICENSE | Pretrained checkpoint license |
The full pipeline uses first-frame Init, temporal Refine in 16-frame windows and Solver in 32-frame windows with a 4-frame overlap and 10 optimization steps. Refine-only inference is also supported.
The checkpoints contain inference network parameters and buffers. SMPL-H/MANO models and DPoser assets must be prepared separately; they are not included in these checkpoint files.
v-dataset checkpoints
The v-dataset pair and matching configs/inference_dataset.yaml provide the pipeline used to generate HKMALA-Motion. It uses first-frame Init, temporal Refine and the full Solver; betas_optimize_windows is 10. The v1 release remains available with its own configuration. Both v1 and v-dataset can run inference on your own marker sequences; try both with their matching configurations and choose the version that works better for your data.
python tools/fetch_checkpoints.py --version v-dataset
python -m src.infer --config configs/inference_dataset.yaml --input my_markers.npz --output outputs/motion.npz --gpu 0
See the version description and inference guide. Model weights use the pretrained checkpoint license. HKMALA-Motion numerical recordings have their own dataset license and access requirements.
Usage
Prepare the environment, body model and checkpoints with Install.md. Run this command from the DirtyMoCap code repository root:
python -m src.infer \
--input demo_samples/sample01.npz \
--output outputs/sample01.npz \
--gpu 0
The included marker-only demos are demo_samples/sample01.npz, demo_samples/sample02.npz and demo_samples/sample03.npz.
See Inference.md, Visualization.md, Evaluation.md and License.md for input/output formats, the viewer, benchmark evaluation and license scope.