Research map / 3D vision / Depth and stereo
Fisheye3R: Adapting Unified 3D Feed-Forward Foundation Models to Fisheye Lenses
Ruxiao Duan, Erin Hong, Dongxu Zhao, Eric Turner, Alex Wong, Yunwen Zhou
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 calibration tokens trained with self-supervised learning on unlabeled perspective images | MapAnything · KITTI360 fisheye | Chamfer distance: 1.301 -> 1.015 | Improved |
| add annotated fisheye data (SL+) to training | MapAnything · KITTI360 fisheye | Chamfer distance: 1.016 -> 0.575 | Improved |
| add supervised perspective labels (SL) over SSL | MapAnything · ScanNet++ fisheye | Chamfer distance: 0.054 -> 0.054 | No clear effect |
| apply Fisheye3R SL+ to VGGT | VGGT · ADT fisheye | Chamfer distance: 0.208 -> 0.059 | Improved |
| add masked attention to calibration tokens (ablation) | Fisheye3R · hybrid ScanNet++ sequences | accuracy across perspective ratios | Improved |
| train tokens on fully fisheye sequences without masked attention (ablation) | Fisheye3R · hybrid ScanNet++ sequences | perspective accuracy | Hurt |
| train tokens on mixed sequences without masked attention (ablation) | Fisheye3R · hybrid ScanNet++ sequences | fisheye accuracy | Hurt |
| insert calibration tokens K=1 instead of 8 (ablation) | Fisheye3R · ScanNet++ MapAnything | Chamfer distance | No clear effect |
| calibrate all 24 encoder layers instead of the last 12 (ablation) | Fisheye3R · ScanNet++ MapAnything | depth Rel: 0.181 vs 0.180 | No clear effect |
| use calibration tokens instead of LoRA r=32 (ablation) | MapAnything · ScanNet++ fisheye | pose/FoV vs depth/point: RPEt 0.134 vs 0.137; Rel 0.171 vs 0.170 | Mixed |
| add masked attention at inference (ablation) | Fisheye3R · MapAnything speed | FPS: 31.1 -> 29.1 | Hurt |
| apply tokens trained with Kannala-Brandt to other projection models | pi3 · Stanford2D3DS | Chamfer distance: 0.264 -> 0.095 (MEI) | Improved |
Closest papers by meaning
- From Perspective to Fisheye Depth Estimation and Open-Vocabulary Segmentation (ECCV 2026 · similarity 0.91)
- Extending Foundational Monocular Depth Estimators to Fisheye Cameras with Calibration Tokens (ICCV 2025 · similarity 0.91)
- Fin3R: Fine-tuning Feed-forward 3D Reconstruction Models via Monocular Knowledge Distillation (NeurIPS 2025 · similarity 0.87)
- PF3plat: Pose-Free Feed-Forward 3D Gaussian Splatting for Novel View Synthesis (ICML 2025 · similarity 0.86)
- Deep Single Image Camera Calibration by Heatmap Regression to Recover Fisheye Images Under Manhattan World Assumption (CVPR 2024 · similarity 0.86)
- Towards Generalizable Multi-Camera 3D Object Detection via Perspective Rendering (AAAI 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.
