Research map / 3D vision / Depth and stereo
OAMaskFlow: Occlusion-Aware Motion Mask for Scene Flow
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes OAMaskFlow, which supervises RAFT-3D's motion embedding with an occlusion-aware motion (OAM) mask derived from photometric/depth consistency and adds attention-based 3D motion propagation to occluded regions, also applying the mask to DROID-SLAM.
Problem
Implicitly learned motion embeddings in end-to-end scene flow networks fail to group pixels of the same rigid object in occluded or inconsistent regions, degrading 3D motion accuracy.
Approach
Generate a ground-truth occlusion-aware motion mask from photometric and geometric consistency (static, occluded, and per-instance labels), use it to explicitly supervise the dense motion embedding, and propagate reliable SE3 motion into occluded regions via self-attention.
Key results
- On FlyingThings3D, EPE3D drops from 0.062 (RAFT-3D) to 0.049 (21.0% reduction) with 7.7M parameters versus 45.8M.
- On KITTI scene flow, all-pixel scene flow error falls from 5.77% (RAFT-3D) to 4.37%; with background refinement it reaches 4.21% on background.
- Applied to DROID-SLAM, average monocular TartanAir ATE falls from 0.35 m to 0.12 m (65.7% reduction).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add OAM mask supervision and motion propagation | RAFT-3D · FlyingThings3D | EPE3D: 21.0% | Helped |
| Add OAM mask and motion propagation | RAFT-3D · KITTI 2015 scene flow benchmark, all pixels | Scene Flow error: 5.77% to 4.37% | Helped |
| Apply OAM mask in DROID-SLAM | DROID-SLAM · TartanAir monocular SLAM challenge | ATE (m): 65.7% | Helped |
| Apply OAM mask in DROID-SLAM | DROID-SLAM · TartanAir stereo SLAM challenge | ATE (m): 58.3% | Helped |
| Add OAM mask supervision (ablation) | RAFT-3D baseline without Laplacian · FlyingThings3D | EPE2D: 2.39 -> 2.01 | Helped |
| Add OAM mask supervision (ablation) | RAFT-3D baseline without Laplacian · FlyingThings3D | EPE3D: 0.062 -> 0.054 | Helped |
| Add motion propagation to OAM mask supervision (ablation) | RAFT-3D baseline without Laplacian · FlyingThings3D | EPE2D: 2.01 -> 1.92 | Helped |
| Add motion propagation to OAM mask supervision (ablation) | RAFT-3D baseline without Laplacian · FlyingThings3D | EPE3D: 0.054 -> 0.052 | Helped |
Surprising findings
- The OAM-supervised model has 7.7M parameters, a 83.2% reduction from RAFT-3D's 45.8M, obtained by shrinking the context encoder.
Limitations
- Fully occluded objects cannot be accurately estimated due to insufficient measurements.
- Depth is only concatenated with RGB as input, underusing 3D structural information.
- KITTI finetuning freezes the motion embedding branch because no camera pose or accurate depth/instance labels are available.
Experimental setup
Datasets: FlyingThings3D KITTI scene flow 2015 TartanAir
Benchmarks: FlyingThings3D KITTI scene flow 2015 TartanAir SLAM challenge
Metrics: EPE2D EPE3D ACC1px ACC.05m ACC.1m D1/D2/optical flow/scene flow error % ATE
Baselines: RAFT-3D CamLiRAFT DELFlow CamLiFlow M-FUSE RigidMask DROID-SLAM DROID-VO ORB-SLAM TartanVO
Builds on: RAFT-3D DROID-SLAM GMFlow
Closest papers by meaning
- FlowMamba: Learning Point Cloud Scene Flow with Global Motion Propagation (AAAI 2025 · similarity 0.86)
- EA3D: Online Open-World 3D Object Extraction from Streaming Videos (NeurIPS 2025 · similarity 0.85)
- OCAI: Improving Optical Flow Estimation by Occlusion and Consistency Aware Interpolation (CVPR 2024 · similarity 0.85)
- STCOcc: Sparse Spatial-Temporal Cascade Renovation for 3D Occupancy and Scene Flow Prediction (CVPR 2025 · similarity 0.84)
- DifFlow3D: Toward Robust Uncertainty-Aware Scene Flow Estimation with Iterative Diffusion-Based Refinement (CVPR 2024 · similarity 0.84)
- VoxelSplat: Dynamic Gaussian Splatting as an Effective Loss for Occupancy and Flow Prediction (CVPR 2025 · similarity 0.84)
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.