Research map / Architectures and efficiency / RNNs and normalization
Selective-Stereo: Adaptive Frequency Information Selection 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 Selective Recurrent Unit (SRU) with a Contextual Spatial Attention module that adaptively fuses multi-frequency (multi-kernel GRU) hidden disparity states inside iterative stereo networks, yielding Selective-RAFT and Selective-IGEV.
Problem
Iterative stereo matching methods use fixed receptive fields in their recurrent update units, so they struggle to capture high-frequency edge details and low-frequency textureless regions simultaneously, causing blurred edges and false matches.
Approach
Replace the single GRU update with parallel GRU branches of small and large kernel sizes whose hidden states are fused by attention maps derived from the context features (CSA), so each image region adaptively selects the frequency band it needs; the module is plugged into RAFT-Stereo, IGEV-Stereo and DLNR.
Key results
- Selective-IGEV reaches a new state-of-the-art EPE of 0.44 on Scene Flow, versus 0.47 for IGEV-Stereo.
- Selective-RAFT reaches EPE 0.47 on Scene Flow vs 0.53 for RAFT-Stereo, with parameters rising only from 11.12M to 11.65M.
- Selective-IGEV ranks 1st among published methods on KITTI 2012, KITTI 2015, ETH3D and Middlebury leaderboards in most metrics.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add multi-kernel SRUs in place of RAFT-Stereo GRUs (ablation) | RAFT-Stereo · Scene Flow test | EPE (px): 0.53 -> 0.50 | Helped |
| invert the CSA attention-map weights (ablation) | RAFT-Stereo with SRUs · Scene Flow test | EPE (px): 0.47 -> 0.50 | Hurt |
| add normal CSA attention weights to SRUs (ablation) | RAFT-Stereo · Scene Flow test | EPE (px): 0.50 -> 0.47 | Helped |
| replace GRUs with SRUs (ablation) | IGEV-Stereo · Scene Flow test | EPE (px): 0.47 -> 0.44 | Helped |
| replace GRUs inside SRUs with LSTMs (ablation) | DLNR · Scene Flow test | EPE (px): 0.49 -> 0.46 | Helped |
| use 8 update iterations instead of 32 (ablation) | Selective-RAFT versus RAFT-Stereo · Scene Flow test | EPE (px): 0.58 -> 0.53 | Helped |
| use 8 update iterations instead of 32 (ablation) | Selective-IGEV versus IGEV-Stereo · Scene Flow test | EPE (px): 0.50 -> 0.48 | Helped |
| use kernel sizes 1 × 1 and 1 × 3 instead of 1 × 1 and 1 × 5 (ablation) | Selective-RAFT · Scene Flow test | EPE (px): 0.48 -> 0.47 | Helped |
| replace IGEV-Stereo's update units with SRUs | IGEV-Stereo · KITTI 2012 test | 2-noc: 1.71 -> 1.59 | Helped |
| replace IGEV-Stereo's update units with SRUs | IGEV-Stereo · KITTI 2015 test | D1-all: 1.59 -> 1.55 | Helped |
| replace IGEV-Stereo's update units with SRUs | IGEV-Stereo · ETH3D test | Bad 0.5: 3.52 -> 3.06 | Helped |
| replace IGEV-Stereo's update units with SRUs | IGEV-Stereo · Middlebury test | Bad 2.0: 4.83 -> 2.51 | Helped |
Surprising findings
- Inverting the CSA attention weights lowers accuracy (Scene Flow EPE 0.50 vs 0.47), confirming the attention maps carry meaningful frequency selection.
Limitations
- SRU receptive fields are still limited by predefined kernel sizes.
- Adding branches or larger kernels increases memory and time costs.
- Gains over IGEV-Stereo on Scene Flow edge regions are only slight because of its aggregated cost volume.
Experimental setup
Datasets: Scene Flow KITTI 2012 KITTI 2015 ETH3D Middlebury Tartan Air CREStereo Dataset Sintel Stereo InStereo2k Falling Things CARLA HR-VS
Benchmarks: Scene Flow KITTI 2012 KITTI 2015 ETH3D Middlebury
Metrics: EPE >1px error 2-noc 2-all D1-all Bad 0.5 Bad 1.0 Bad 2.0 AvgErr run-time
Baselines: RAFT-Stereo IGEV-Stereo DLNR CREStereo LaC + GANet ACVNet LEAStereo CroCo-Stereo GMStereo HITNet
Builds on: RAFT-Stereo IGEV-Stereo DLNR CBAM octave convolution
Closest papers by meaning
- Any-Stereo: Arbitrary Scale Disparity Estimation for Iterative Stereo Matching (AAAI 2024 · similarity 0.87)
- DEFOM-Stereo: Depth Foundation Model Based Stereo Matching (CVPR 2025 · similarity 0.86)
- All-in-One: Transferring Vision Foundation Models into Stereo Matching (AAAI 2025 · similarity 0.86)
- Hierarchical Object-Aware Dual-Level Contrastive Learning for Domain Generalized Stereo Matching (NeurIPS 2024 · similarity 0.85)
- Stereo Anywhere: Robust Zero-Shot Deep Stereo Matching Even Where Either Stereo or Mono Fail (CVPR 2025 · similarity 0.85)
- Consistency-aware Self-Training for Iterative-based Stereo Matching (CVPR 2025 · 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.