Research map / 3D vision / Depth and stereo
Selfi: Self-improving Reconstruction Engine via 3D Geometric Feature Alignment
Youming Deng, Songyou Peng, Junyi Zhang, Kathryn Heal, Tiancheng Sun, John Flynn, Steve Marschner, Lucy Chai
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 geometric feature alignment (vs VGGT tracking-head features) (ablation) | Selfi Gaussian head · DL3DV ablation | PSNR: 22.53 -> 23.29 | Improved |
| add view-dependent density SH on top of RGB SH (ablation) | Selfi · DL3DV ablation | PSNR: 23.70 -> 24.67 | Improved |
| add RGB SH (without density SH) (ablation) | Selfi · DL3DV ablation | PSNR: 23.29 -> 23.70 | Improved |
| swap in BA-refined poses without depth shift (ablation) | Selfi · DL3DV ablation | PSNR: 24.67 -> 24.61 | Hurt |
| add affine depth shift after BA (ablation) | Selfi · DL3DV ablation | PSNR: 24.61 -> 24.88 | Improved |
| apply BA with aligned features and depth shift (ablation) | Selfi · DL3DV | PSNR: 24.94 -> 25.13 | Improved |
| apply BA with aligned features | Selfi · DL3DV | AUC@3: 0.813 -> 0.870 | Improved |
| apply BA with aligned features, zero-shot | Selfi · Tanks&Temples | AUC@3: 0.818 -> 0.955 | Improved |
| use aligned-feature matching instead of Co-Tracker for BA | VGGT poses · Pose estimation (10 frames) | AUC@3: 0.8350 -> 0.8670 | Improved |
| use aligned-feature matching instead of Co-Tracker for BA | VGGT poses · Pose estimation (100 frames) | AUC@3: Out-of-Memory -> 0.8762 | Improved |
| Selfi vs 3DGS with GT poses | Selfi · RealEstate10K (6 frames) | PSNR: 28.46 -> 28.34 | No clear effect |
| Selfi vs pose-required methods, two-view | Selfi · RealEstate10K (two-view) | PSNR: 29.72 -> 29.01 | Mixed |
Closest papers by meaning
- LoRA3D: Low-Rank Self-Calibration of 3D Geometric Foundation models (ICLR 2025 · similarity 0.88)
- Less Decoder is More Encoder: Geometric Representation Learning from Novel View Synthesis (NeurIPS 2026 · similarity 0.87)
- Sparse-view Pose Estimation and Reconstruction via Analysis by Generative Synthesis (NeurIPS 2024 · similarity 0.87)
- Fin3R: Fine-tuning Feed-forward 3D Reconstruction Models via Monocular Knowledge Distillation (NeurIPS 2025 · similarity 0.87)
- VGGT: Visual Geometry Grounded Transformer (CVPR 2025 · similarity 0.87)
- PF3plat: Pose-Free Feed-Forward 3D Gaussian Splatting for Novel View Synthesis (ICML 2025 · 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.
