Research map / 3D vision / NeRF and radiance fields
Learning 3D-aware GANs from Unposed Images with Template Feature Field
Xinya Chen, Hanlei Guo, Yanrui Bin, Shangzhan Zhang, Yuanbo Yang, Yujun Shen, Yue Wang, Yiyi Liao
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 |
|---|---|---|---|
| estimate poses with a learned template feature field instead of letting the generator learn the pose distribution (vs 3DGP/PoF3D) | 3D-aware GAN (EG3D-based) · CompCars | Depth_360: 0.95 (EG3D), 4.02 (3DGP), 10.31 (PoF3D) -> 0.31 | Improved |
| use template feature field rather than template radiance field (RGB) for pose estimation (ablation) | TeFF pose estimation · ShapeNet Cars | elevation KL divergence: 0.1422 -> 0.0696 | Improved |
| use template feature field rather than gray-image template radiance field (ablation) | TeFF pose estimation · ShapeNet Cars | azimuth KL divergence: 0.0656 -> 0.0555 | Improved |
| add phase correlation to model scale and in-plane rotation (4 DoF instead of 2 DoF) (ablation) | TeFF · CompCars | Depth_360: 4.98 -> 0.31 | Improved |
| add phase correlation (4 DoF instead of 2 DoF) (ablation) | TeFF · CompCars | FID_est: 11.09 -> 20.60 | Hurt |
| increase azimuth x elevation bins from 12x6 to 36x18 (ablation) | TeFF · ShapeNet Cars | FID_gt: 11.29 -> 5.95 | Improved |
| increase bins beyond 36x18 (to 60x30) (ablation) | TeFF · ShapeNet Cars | FID_gt: 5.95 -> 5.96 | No clear effect |
| use phase correlation instead of naive grid search over scale and rotation (ablation) | TeFF · runtime (A6000) | early-stage template rendering time (s/iter): 4423.0519 -> 0.0992 | Improved |
| estimate pose with TeFF versus EG3D with uniform pose sampling | TeFF · CompCars | FID_360: 7.06 -> 27.71 | Hurt |
| TeFF pose estimation versus jointly learned pose distribution | TeFF · ShapeNet Cars | azimuth KL divergence: 4.4829 (PoF3D) -> 0.0555 | Improved |
| apply TeFF to face/cat data without GT poses | TeFF · FFHQ | FID_gt: 4.80 (EG3D) -> 5.64 | Mixed |
Closest papers by meaning
- Pix2NeRF: Unsupervised Conditional p-GAN for Single Image to Neural Radiance Fields Translation (CVPR 2022 · similarity 0.88)
- Learning Generative Models of Textured 3D Meshes From Real-World Images (ICCV 2021 · similarity 0.88)
- Shape, Pose, and Appearance From a Single Image via Bootstrapped Radiance Field Inversion (CVPR 2023 · similarity 0.88)
- Learning 3D-Aware Image Synthesis With Unknown Pose Distribution (CVPR 2023 · similarity 0.88)
- LOLNerf: Learn From One Look (CVPR 2022 · similarity 0.87)
- HoloGAN: Unsupervised Learning of 3D Representations From Natural Images (ICCV 2019 · 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.
