Research map / Video / Video generation
Action-Conditioned 3D Human Motion Synthesis With Transformer VAE
Mathis Petrovich, Michael J. Black, Gül Varol
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| combine rotation loss with vertex loss (ablation) | ACTOR · UESTC | accuracy (%): 42.4 -> 91.1 | Improved |
| add vertex loss to rotation loss (ablation) | ACTOR · NTU-13 | FID_tr: 0.23 -> 0.19 | Improved |
| use coordinate-based losses (joint or vertex) alone (ablation) | ACTOR · UESTC | accuracy (%): 3.3 and 2.7 versus 91.1 | Hurt |
| replace GRU with Transformer encoder-decoder (ablation) | ACTOR · UESTC | FID_test: 27.08 -> 23.43 | Improved |
| replace fully connected with Transformer (ablation) | ACTOR · UESTC | FID_test: 548.13 -> 23.43 | Improved |
| add one-shot decoding with positional encodings instead of autoregressive decoder (ablation) | ACTOR · UESTC | FID_test: 60.10 -> 23.43 | Improved |
| add distribution parameter tokens instead of averaging encoder outputs (ablation) | ACTOR · UESTC | accuracy (%): 86.2 -> 91.1 | Improved |
| add additive action bias token instead of one-hot concatenation (ablation) | ACTOR · UESTC | accuracy (%): 89.4 -> 91.1 | Improved |
| add additive action bias token instead of one-hot concatenation (ablation) | ACTOR · NTU-13 | FID_tr: 0.16 -> 0.19 | Mixed |
| set lambda_KL to 1e-5 rather than other values (ablation) | ACTOR · UESTC | accuracy (%): 90.5 vs 34.4 at 1e-3, 77.4 at 1e-6 | Improved |
| use batch size 20 instead of 10 (ablation) | ACTOR · UESTC | accuracy (%): 39.7 -> 90.5 | Improved |
| increase Transformer layers from 2 to 8 (ablation) | ACTOR · UESTC | accuracy (%): 84.9 -> 90.5 | Improved |
Closest papers by meaning
- MT-VAE: Learning Motion Transformations to Generate Multimodal Human Dynamics (ECCV 2018 · similarity 0.91)
- Learning Uncoupled-Modulation CVAE for 3D Action-Conditioned Human Motion Synthesis (ECCV 2022 · similarity 0.91)
- A Unified 3D Human Motion Synthesis Model via Conditional Variational Auto-Encoder (ICCV 2021 · similarity 0.90)
- Conditional Temporal Variational AutoEncoder for Action Video Prediction (IJCV 2023 · similarity 0.90)
- A Unified Masked Autoencoder with Patchified Skeletons for Motion Synthesis (AAAI 2024 · similarity 0.88)
- MoMask: Generative Masked Modeling of 3D Human Motions (CVPR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
