Research map / Label-efficient and robust learning / Self-supervised and contrastive
E-RayZer: Self-supervised 3D Reconstruction as Spatial Visual Pre-training
Qitao Zhao, Hao Tan, Qianqian Wang, Sai Bi, Kai Zhang, Kalyan Sunkavalli, Shubham Tulsiani, Hanwen Jiang
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 |
|---|---|---|---|
| use explicit 3D Gaussians instead of RayZer's implicit latent scene | E-RayZer vs RayZer · DL3DV pose estimation, DL3DV training | RPA@5: 0.0 -> 72.0 | Improved |
| use explicit 3D Gaussians instead of implicit latent scene | E-RayZer vs RayZer · WildRGB-D novel-view synthesis, DL3DV training | PSNR: 25.9 -> 24.3 | Hurt |
| add visual-overlap curriculum (geometric) versus no curriculum (ablation) | E-RayZer · DL3DV training, DL3DV pose | RPA@5: 4.0 -> 72.0 | Improved |
| visual-overlap curriculum instead of frame-interval curriculum (ablation) | E-RayZer · DL3DV training, DL3DV pose | RPA@5: 56.1 -> 72.0 | Improved |
| semantic overlap instead of geometric overlap for curriculum (ablation) | E-RayZer · DL3DV and 7-dataset training | RPA@5: 73.2 vs 72.0 (DL3DV); 58.7 vs 59.9 (7-dataset) | No clear effect |
| initialize VGGT* with E-RayZer weights | VGGT* · DL3DV in-domain pose | RPA@5: 79.6 -> 87.3 | Improved |
| train on 7-dataset mix instead of DL3DV only (ablation) | E-RayZer · CO3Dv2 pose | RPA@5: 19.1 -> 30.3 | Improved |
| train on 7-dataset mix instead of DL3DV only (ablation) | E-RayZer · DL3DV pose (in-domain) | RPA@5: 72.0 -> 59.9 | Hurt |
| train on DL3DV instead of RE10K (ablation) | E-RayZer · DL3DV pose | RPA@5: 21.2 -> 72.0 | Improved |
| use E-RayZer encoder instead of DINOv2/DINOv3/CroCo v2/VideoMAE V2 (frozen) | multi-view depth and pose probing · ScanNet++ | AbsRel: 0.193 (DINOv2) -> 0.116 | Improved |
| use E-RayZer encoder instead of RayZer for flow | pairwise flow probing · StaticThings3D | EPE: 1.105 -> 1.254 | Hurt |
| E-RayZer versus supervised VGGT* when trained on large data | E-RayZer · 7-dataset training, DL3DV pose | RPA@5: 66.1 (VGGT*) vs 59.9 (E-RayZer) | Hurt |
Closest papers by meaning
- RayZer: A Self-supervised Large View Synthesis Model (ICCV 2025 · similarity 0.92)
- RayDer: Scalable Self-Supervised Novel View Synthesis from Real-World Video (ECCV 2026 · similarity 0.88)
- From None to All: Self-Supervised 3D Reconstruction via Novel View Synthesis (CVPR 2026 · similarity 0.87)
- Epipolar-Free 3D Gaussian Splatting for Generalizable Novel View Synthesis (NeurIPS 2024 · similarity 0.85)
- DistillNeRF: Perceiving 3D Scenes from Single-Glance Images by Distilling Neural Fields and Foundation Model Features (NeurIPS 2024 · similarity 0.85)
- RaySt3R: Predicting Novel Depth Maps for Zero-Shot Object Completion (NeurIPS 2025 · similarity 0.85)
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.
