Research map / 3D vision / Depth and stereo
GRF-Recon: Global Ray-Field Optimization for Long-Sequence Feed-forward Reconstruction
Enpeng Li, Yunzhou Zhang, Zhiyao Zhang, Dexuan Lyu, Chenyu Wang, Chiyuan Cui, Cheng Cheng
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 re-normalized LoRA distillation of monocular priors (ablation) | DA3 front-end · KITTI seq 00 | ATE RMSE (m): 6.80 -> 4.35 | Improved |
| add re-normalized LoRA fine-tuning to DA3 | DA3 · KITTI Depth | Rel: 5.14 -> 4.50 | Improved |
| add hybrid ray-matching module (ablation) | GRF-Recon · KITTI seq 00 | ATE RMSE (m): 7.35 -> 4.35 | Improved |
| add loop closure (ablation) | GRF-Recon · KITTI seq 00 | ATE RMSE (m): 8.12 -> 4.35 | Improved |
| add global backend optimization (ablation) | GRF-Recon · KITTI seq 00 | ATE RMSE (m): 12.35 -> 4.35 | Improved |
| use all three ray-matching terms (ray, gradient, photometric) instead of two (ablation) | GRF-Recon · KITTI seq 07 | ATE RMSE (m): 2.45 -> 1.64 (vs ray+grad only) | Improved |
| change chunk size/overlap among 30/10, 60/10, 60/20 (ablation) | GRF-Recon · KITTI seq 00 | ATE RMSE (m): 4.41 / 4.35 / 4.32 | No clear effect |
| use GRF-Recon instead of DA3-Streaming for camera tracking | long-sequence feed-forward reconstruction · KITTI (avg excl. seq 01) | ATE RMSE (m): 12.72 -> 7.18 | Improved |
| use GRF-Recon instead of DA3-Streaming | long-sequence feed-forward reconstruction · Waymo Open | ATE RMSE (m): 1.76 -> 1.39 | Improved |
| use GRF-Recon vs VGGT-Long on Waymo Chamfer | long-sequence reconstruction · Waymo reconstruction seq 460 | Chamfer: 1.420 (VGGT-Long) vs 1.480 | Hurt |
| GRF-Recon vs calibrated ORB-SLAM3 | tracking · KITTI seq 01 | ATE RMSE (m): 7.23 (ORB-SLAM3) vs 76.52 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DROID-SLAM — DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras (NeurIPS 2021) · better on 1 of 1 shared result
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
- Raymap-Guided Coupling for Drift-Robust Unposed Feed-Forward 3D Reconstruction (ECCV 2026 · similarity 0.90)
- Rethinking the State Update Gate for Long-Sequence Recurrent 3D Reconstruction (NeurIPS 2026 · similarity 0.88)
- RayMap3R: Inference-Time RayMap for Dynamic 3D Reconstruction (ECCV 2026 · similarity 0.87)
- GO-SLAM: Global Optimization for Consistent 3D Instant Reconstruction (ICCV 2023 · similarity 0.87)
- CAM3R: Camera-Agnostic Model for 3D Reconstruction (ECCV 2026 · similarity 0.87)
- ReCon-GS: Continuum-Preserved Guassian Streaming for Fast and Compact Reconstruction of Dynamic Scenes (NeurIPS 2025 · similarity 0.86)
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.
