Research map / 3D vision / Depth and stereo
MoCha-Stereo: Motif Channel Attention Network for Stereo Matching
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 MoCha-Stereo, a stereo matching network that restores lost edge geometry in feature channels via motif channel attention, a motif channel correlation volume, and a reconstruction-error motif penalty refinement module.
Problem
Learning-based stereo matching loses geometric structure in feature channels during channel generation, causing edge detail mismatches in disparity maps.
Approach
Mines repeated geometric contours (motifs) across feature channels using adaptive sliding windows in the frequency domain, projects them onto normal channels via a CAMP-guided correlation volume, and separates low- and high-frequency reconstruction-error components in a REMP refinement module.
Key results
- Achieves EPE 0.41 on Scene Flow, 12.77% better than IGEV-Stereo (0.47).
- Ranks 1st on KITTI-2012 and KITTI-2015 leaderboards; KITTI-2015 all-pixel D1 1.53 vs 1.59 for IGEV-Stereo.
- Ablation: MCCV reduces EPE from 0.458 to 0.423 and REMP to 0.445; full model reaches 0.412.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add MoCha-Stereo's motif-channel matching framework | Stereo matching · Scene Flow test | EPE (px): 0.47 -> 0.41 | Helped |
| use MoCha-Stereo instead of IGEV-Stereo | Stereo matching · KITTI-2012 All | Out-Noc (%): 1.12 -> 1.06 | Helped |
| use MoCha-Stereo instead of IGEV-Stereo | Stereo matching · KITTI-2012 Reflective | Out-Noc (%): 1.0 -> 0.8 | Helped |
| use MoCha-Stereo instead of IGEV-Stereo | Stereo matching · KITTI-2015 All pixels | D1-error all (%): 1.59 -> 1.53 | Helped |
| use MoCha-Stereo instead of IGEV-Stereo | Stereo matching · Middlebury Half | 2-pixel error rate: 7.1 -> 6.2 | Helped |
| use MoCha-Stereo instead of IGEV-Stereo | Stereo matching · Middlebury Quarter | 2-pixel error rate: 6.2 -> 4.9 | Helped |
| use MoCha-Stereo instead of IGEV-Stereo | Stereo matching · ETH3D | 1-pixel error rate: 3.6 -> 3.2 | Helped |
| add MCA to feature maps (ablation) | IGEV-Stereo with EfficientNet · Scene Flow test ablation | EPE (px): 0.458 -> 0.449 | Helped |
| add REMP (ablation) | IGEV-Stereo with EfficientNet · Scene Flow test ablation | EPE (px): 0.458 -> 0.445 | Helped |
| add MCA and REMP (ablation) | IGEV-Stereo with EfficientNet · Scene Flow test ablation | EPE (px): 0.458 -> 0.438 | Helped |
| add MCCV (ablation) | IGEV-Stereo with EfficientNet · Scene Flow test ablation | EPE (px): 0.458 -> 0.423 | Helped |
| add MCA, REMP, and MCCV (ablation) | IGEV-Stereo with EfficientNet · Scene Flow test ablation | EPE (px): 0.458 -> 0.412 | Helped |
Limitations
- Full model adds parameters (20.74M vs 18.31M baseline) and per-iteration cost is higher than baseline in ablations.
- Motif windows are learned parameters rather than mined from data, with the motivation resting on heuristic arguments.
Experimental setup
Datasets: Scene Flow
Benchmarks: KITTI 2012 KITTI 2015 Scene Flow ETH3D Middlebury DTU
Metrics: EPE D1-error Out-Noc Avg-Noc accuracy completeness overall
Baselines: IGEV-Stereo DLNR UPFNet ACVNet RAFT-Stereo CREStereo Lac-GwcNet CroCo-Stereo
Builds on: IGEV-Stereo RAFT-Stereo GwcNet DLNR Matrix Profile
Closest papers by meaning
- Attention Concatenation Volume for Accurate and Efficient Stereo Matching (CVPR 2022 · similarity 0.86)
- Geometry-Aware Stereo Matching via Monocular Disparity Distribution Prior and Gradient Enhancement (AAAI 2026 · similarity 0.86)
- Revisiting Domain Generalized Stereo Matching Networks From a Feature Consistency Perspective (CVPR 2022 · similarity 0.86)
- CFNet: Cascade and Fused Cost Volume for Robust Stereo Matching (CVPR 2021 · similarity 0.85)
- Selective-Stereo: Adaptive Frequency Information Selection for Stereo Matching (CVPR 2024 · similarity 0.85)
- Local Similarity Pattern and Cost Self-Reassembling for Deep Stereo Matching Networks (AAAI 2022 · 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.