Research map / 3D vision / Human pose estimation
Human Motion Transfer With 3D Constraints and Detail Enhancement
Yang-Tian Sun, Qiancheng Fu, Yue-Ren Jiang, Zitao Liu, Yu‐Kun Lai, Hongbo Fu, Lin Gao
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 |
|---|---|---|---|
| add 3D projection condition to 2D pose condition (MT(2D+3D)) (ablation) | MT-Net · Dance dataset, cross-subject transfer | FID: 58.62 -> 55.77 | Improved |
| add 3D projection condition to 2D pose condition (MT(2D+3D)) (ablation) | MT-Net · Dance dataset, self-transfer | SSIM: 0.828 -> 0.856 | Improved |
| use 3D condition alone instead of 2D condition (ablation) | MT-Net · Dance dataset, cross-subject transfer | IS-ReID: 3.523 -> 3.632 | Improved |
| add DE-Net detail enhancement to MT(2D+3D) (ablation) | full pipeline · Dance dataset, self-transfer | SSIM: 0.856 -> 0.891 | Improved |
| add DE-Net to MT(2D) (ablation) | MT-Net · Dance dataset, self-transfer | LPIPS: 0.064 -> 0.043 | Improved |
| add DE-Net to MT(2D) (ablation) | MT-Net · Dance dataset, cross-subject transfer | IS-ReID: 3.523 -> 3.587 | Improved |
| use learned blending parameter instead of average blending (ablation) | DE-Net · Dance dataset, cross-subject transfer | FID: 53.76 -> 51.26 | Improved |
| use learned blending parameter instead of average blending (ablation) | DE-Net · Dance dataset, self-transfer | SSIM: 0.887 -> 0.891 | Improved |
| use Laplacian-projection 3D pose labels instead of UV-coordinate pose labels of prior work (ablation) | pose GAN condition · Dance dataset, cross-subject transfer | FID: 58.42 -> 51.26 | Improved |
| use Laplacian-projection 3D pose labels instead of prior pose labels (ablation) | pose GAN condition · Dance dataset, cross-subject transfer | IS-ReID: 3.612 -> 4.015 | Improved |
| add 3D condition to MT(2D)+DE (self-transfer, same body shape) (ablation) | full pipeline · Dance dataset, self-transfer | SSIM: 0.877 -> 0.891 | No clear effect |
| compare to vid2vid, Chan et al. and Liquid Warping GAN | Ours · Dance dataset, cross-subject transfer | FID: 57.02 -> 51.26 | Improved |
Closest papers by meaning
- Human Action Transfer Based on 3D Model Reconstruction (AAAI 2019 · similarity 0.91)
- Scene-Aware Generative Network for Human Motion Synthesis (CVPR 2021 · similarity 0.88)
- Pose Guided Human Video Generation (ECCV 2018 · similarity 0.88)
- Robust Pose Transfer With Dynamic Details Using Neural Video Rendering (TPAMI 2022 · similarity 0.87)
- High-Fidelity Neural Human Motion Transfer From Monocular Video (CVPR 2021 · similarity 0.87)
- ActFormer: A GAN-based Transformer towards General Action-Conditioned 3D Human Motion Generation (ICCV 2023 · similarity 0.87)
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.
