Research map / 3D vision / Depth and stereo
FlowSeek: Optical Flow Made Easier with Depth Foundation Models and Motion Bases
Matteo Poggi, Fabio Tosi
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 BaseNet with motion bases from depth foundation model (ablation) | SEA-RAFT (S) · TartanAir val (100K-step ablation) | EPE: 1.38 -> 1.04 | Improved |
| add depth features to the BaseNet configuration (ablation) | FlowSeek (T) · KITTI 2012 train (TartanAir-trained ablation) | Fl-All: 4.23 -> 4.16 | Improved |
| add depth input to ContextNet on top of BaseNet and features (ablation) | FlowSeek (T) · KITTI 2012 train (TartanAir-trained ablation) | Fl-All: 4.16 -> 4.19 | No clear effect |
| use motion bases instead of raw depth map as BaseNet input (ablation) | FlowSeek (T) · TartanAir val (100K-step ablation) | EPE: 1.05 -> 1.03 | Improved |
| use motion bases instead of depth map after fine-tuning on KITTI 2015 (ablation) | FlowSeek (T) · KITTI 2015 val dynamic regions | Fl-EPE: 2.49 -> 2.60 | Hurt |
| use motion bases instead of depth map after fine-tuning on KITTI 2015 (ablation) | FlowSeek (T) · KITTI 2015 val static regions | Fl-EPE: 1.21 -> 1.06 | Improved |
| scale from small (12 iters) to medium (ResNet-34, Depth Anything v2 B, 4 iters) (ablation) | FlowSeek · TartanAir val (100K-step ablation) | EPE: 0.99 -> 0.90 | Mixed |
| run 12 iterations instead of 4 with ResNet-34 (ablation) | FlowSeek (M) to (L) · TartanAir val (100K-step ablation) | EPE: 0.90 -> 0.85 | Improved |
| use Depth Anything v2 instead of DPT-Hybrid (ablation) | FlowSeek (T) · TartanAir val (100K-step ablation) | EPE: 1.08 -> 1.03 | Improved |
| add FlowSeek priors to CRAFT backbone | CRAFT · TartanAir val (100K-step ablation) | EPE: 1.77 -> 1.39 | Improved |
| add FlowSeek priors to FlowFormer backbone | FlowFormer · TartanAir val (100K-step ablation) | EPE: 1.63 -> 1.30 | Improved |
| train SEA-RAFT (L) on a single GPU instead of original budget | SEA-RAFT (L) · KITTI 2015 train (TartanAir pretrain zero-shot) | Fl-all: 12.7 -> 13.5 | Hurt |
Closest papers by meaning
- FlowNet 2.0: Evolution of Optical Flow Estimation With Deep Networks (CVPR 2017 · similarity 0.89)
- Learning Optical Flow From Still Images (CVPR 2021 · similarity 0.89)
- Learning End-to-End Scene Flow by Distilling Single Tasks Knowledge (AAAI 2020 · similarity 0.88)
- MEMFOF: High-Resolution Training for Memory-Efficient Multi-Frame Optical Flow Estimation (ICCV 2025 · similarity 0.88)
- ADFactory: An Effective Framework for Generalizing Optical Flow with NeRF (CVPR 2024 · similarity 0.87)
- UnSAMFlow: Unsupervised Optical Flow Guided by Segment Anything Model (CVPR 2024 · 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.
