Research map / Label-efficient and robust learning / Self-supervised and contrastive
RAFT-MSF: Self-Supervised Monocular Scene Flow Using Recurrent Optimizer
Bayram Bayramli, Junhwa Hur, Hongtao Lu
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 gradient detaching on the convex upsampling mask head (ablation) | RAFT-MSF · KITTI 2015 Scene Flow Training | SF-all (%): 38.37 -> 37.37 | Improved |
| add disparity initialization with two conv layers on top of gradient detach (ablation) | RAFT-MSF · KITTI 2015 Scene Flow Training | SF-all (%): 37.37 -> 35.62 | Improved |
| add gradient detach and disparity init together to RAFT baseline (ablation) | RAFT-MSF · KITTI 2015 Scene Flow Training | D1-all (%): 26.95 -> 24.04 | Improved |
| initialize disparity with 1 conv layer instead of zeros (ablation) | RAFT-MSF · KITTI 2015 Scene Flow Training | SF-all (%): 37.37 -> 36.15 | Improved |
| use 2 conv layers instead of 1 for disparity initialization (ablation) | RAFT-MSF · KITTI 2015 Scene Flow Training | SF-all (%): 36.15 -> 35.62 | Improved |
| initialize disparity with one or half of maximum disparity instead of zeros (ablation) | RAFT-MSF · KITTI 2015 Scene Flow Training | training convergence | Hurt |
| use convex upsampling without gradient detaching in self-supervised setup (ablation) | RAFT-MSF · KITTI 2015 Scene Flow Training | artifacts | Hurt |
| replace prior two-view method with RAFT-based baseline | RAFT baseline · KITTI 2015 Scene Flow Training | SF-all (%): 47.05 -> 38.27 | Improved |
| RAFT-MSF vs prior multi-view method using only two views | RAFT-MSF · KITTI 2015 Scene Flow Training | SF-all (%): 39.82 -> 30.97 | Improved |
| fine-tune with 200 annotated pairs | RAFT-MSF · KITTI 2015 Scene Flow Test | SF-all (%): 34.98 -> 22.50 | Improved |
| fine-tune with 200 annotated pairs | RAFT-MSF · KITTI 2015 Scene Flow Test | Fl-all (%): 18.37 -> 8.80 | Improved |
| RAFT-MSF vs prior depth methods on Eigen split | RAFT-MSF · Eigen split depth | Abs Rel: 0.093 | No clear effect |
Closest papers by meaning
- Self-Supervised Monocular Scene Flow Estimation (CVPR 2020 · similarity 0.93)
- Self-Supervised Multi-Frame Monocular Scene Flow (CVPR 2021 · similarity 0.92)
- Deep Scene Flow Learning: From 2D Images to 3D Point Clouds (TPAMI 2023 · similarity 0.90)
- Zero-Shot Monocular Scene Flow Estimation in the Wild (CVPR 2025 · similarity 0.89)
- ZeroFlow: Scalable Scene Flow via Distillation (ICLR 2024 · similarity 0.88)
- DeltaFlow: An Efficient Multi-frame Scene Flow Estimation Method (NeurIPS 2025 · similarity 0.88)
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.
