Research map / 3D vision / Face and shape
AutoAvatar: Autoregressive Neural Fields for Dynamic Avatar Modeling
Ziqian Bai, Timur Bagautdinov, Javier Romero, Michael Zollhöfer, Ping Tan, Shunsuke Saito
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 autoregressive shape history (H_t and dH) to pose-only input (ablation) | Pose+dPose variant · DFaust interpolation, rollout 1 | Scan-to-prediction distance (mm): 3.852 -> 1.569 | Improved |
| add autoregressive shape history (ablation) | Pose+dPose variant · DFaust extrapolation, rollout 1 | Scan-to-prediction distance (mm): 3.984 -> 1.150 | Improved |
| add autoregressive shape history (ablation) | Pose+dPose variant · DFaust interpolation, rollout 30 | Scan-to-prediction distance (mm): 4.156 -> 5.255 | Hurt |
| add dPose (pose differences) to pose-only input (ablation) | Pose · DFaust interpolation, rollout 1 | Scan-to-prediction distance (mm): 4.218 -> 3.852 | Improved |
| replace single-frame pose input with TCN over 16 poses (ablation) | Pose · DFaust interpolation, rollout 1 | Scan-to-prediction distance (mm): 4.218 -> 4.068 | Improved |
| encode shape with signed heights on articulated observer points instead of localized latent embeddings (ablation) | L-embed · DFaust interpolation, rollout 1 | Scan-to-prediction distance (mm): 1.784 -> 1.569 | Improved |
| encode shape with signed heights instead of localized latent embeddings (ablation) | L-embed · DFaust extrapolation, rollout 30 | Scan-to-prediction distance (mm): 5.192 -> 4.526 | Improved |
| encode shape with signed heights instead of global latent embedding (ablation) | G-embed · DFaust interpolation, rollout 1 | Scan-to-prediction distance (mm): 2.932 -> 1.569 | Improved |
| encode shape with signed heights instead of global latent embedding (ablation) | G-embed · DFaust interpolation, rollout 30 | Scan-to-prediction distance (mm): 3.793 -> 5.255 | Hurt |
| use AutoAvatar instead of SNARF | SNARF · DFaust interpolation, rollout 1 | Scan-to-prediction distance (mm): 7.428 -> 1.569 | Improved |
| use refined SMPL poses instead of AMASS poses | SNARF · DFaust subject 50002 interpolation, rollout 1 | Scan-to-prediction distance (mm): 7.898 -> 3.982 | Improved |
| use refined SMPL poses instead of AMASS poses | AutoAvatar · DFaust subject 50002 interpolation, rollout 1 | Scan-to-prediction distance (mm): 1.731 -> 1.417 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- Pose — pOSE: Pseudo Object Space Error for Initialization-Free Bundle Adjustment (CVPR 2018) · better on 2 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Dynamic Point Fields (ICCV 2023 · similarity 0.90)
- Animatable Implicit Neural Representations for Creating Realistic Avatars From Videos (TPAMI 2024 · similarity 0.89)
- Pose Modulated Avatars from Video (ICLR 2024 · similarity 0.88)
- Structured Local Radiance Fields for Human Avatar Modeling (CVPR 2022 · similarity 0.88)
- Dynamic Neural Radiance Fields for Monocular 4D Facial Avatar Reconstruction (CVPR 2021 · similarity 0.87)
- COAP: Compositional Articulated Occupancy of People (CVPR 2022 · 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.
