Research map / Label-efficient and robust learning / Semi-supervised learning
Consistency-aware Self-Training for Iterative-based 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 CST-Stereo, a teacher-student self-training framework for iterative stereo matching that weights pseudo-labels with multi-resolution and iterative prediction consistency filters.
Problem
Iterative-based stereo matching models rely heavily on labeled data and generalize poorly to unlabeled real-world data; existing cost-volume-based self-training reliability measures do not transfer to iterative models.
Approach
Estimates pixel-wise pseudo-label reliability from prediction fluctuations across input resolutions and across GRU iterations, then applies these as soft weights in the student loss instead of hard thresholding.
Key results
- On Middlebury, CST-Selective-IGEV reaches EPE 1.32 and D1 7.75 vs. 5.28 and 12.07 for the Selective-IGEV baseline in the in-domain setting (ablation table).
- In domain adaptation on Middlebury, CST-Selective-IGEV achieves F 1.32 and H 7.75 without target labels, versus 5.28 and 12.07 for Selective-IGEV.
- Ablation shows naive self-training without filtering degrades Middlebury EPE to 6.15 from 5.28, while the full soft-filtered method reaches 1.32.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply CST-Stereo to Selective-IGEV | Selective-IGEV · KITTI 2015 leaderboard | D1-all: 1.55 -> 1.50 | Helped |
| Apply CST-Stereo to Selective-IGEV | Selective-IGEV · Middlebury leaderboard | Avgerr: 0.91 -> 0.87 | Helped |
| Apply CST-Stereo to CREStereo | CREStereo · ETH3D leaderboard | bad 1.0: 0.98 -> 1.02 | Mixed |
| Add pseudo-label filtering during self-training (ablation) | Selective-IGEV · Middlebury Eval | EPE: 6.15 -> 2.85 | Helped |
| Add pseudo-label filtering during self-training (ablation) | Selective-IGEV · Middlebury Eval | D1: 11.76 -> 9.17 | Helped |
| Add the full consistency-aware soft filtering design (ablation) | Selective-IGEV self-training · Middlebury Eval | EPE: 2.85 -> 1.32 | Helped |
| Add the full consistency-aware soft filtering design (ablation) | Selective-IGEV self-training · Middlebury Eval | D1: 9.17 -> 7.75 | Helped |
| Use MRPCF filter only (ablation) | Selective-IGEV self-training · Middlebury Eval | EPE: 5.28 -> 1.58 | Helped |
| Use IPCF filter only (ablation) | Selective-IGEV self-training · Middlebury Eval | EPE: 5.28 -> 1.70 | Helped |
| Combine MRPCF and IPCF (ablation) | Selective-IGEV self-training · Middlebury Eval | EPE: 1.58 -> 1.32 | Helped |
| Replace the proposed filter with DDE (ablation) | Selective-IGEV self-training · Middlebury Eval | EPE: 1.32 -> 7.47 | Hurt |
| Replace the proposed filter with LRC (ablation) | Selective-IGEV self-training · Middlebury Eval | EPE: 1.32 -> 4.19 | Hurt |
Surprising findings
- Naive self-training without any pseudo-label filtering degrades performance below the baseline on Middlebury (EPE 6.15 vs 5.28).
- Previous cost-volume-based filters (DDE, LRC) are less effective or harmful when applied to iterative-based models.
Limitations
- Final performance remains bounded by the base model's capacity and architecture, as the authors note.
- Filter thresholds and scale factors (epsilon, tau) are hand-set hyperparameters and the EMA decay is sensitive to its value.
Experimental setup
Datasets: SceneFlow Middlebury 2014 KITTI 2015 ETH3D Instereo2k Holopix50k
Benchmarks: Middlebury KITTI 2015 ETH3D Instereo2k
Metrics: EPE D1 bad 1.0 bad 2.0 bad 4.0 Avgerr F H
Baselines: PSMNet RAFT-Stereo CREStereo EAI-Stereo DLNR GMStereo IGEV-Stereo Selective-IGEV Ada-Selective-IGEV ZOLE-Selective-IGEV DDE LRC
Builds on: RAFT-Stereo Selective-IGEV DINO mean teacher
Closest papers by meaning
- DualNet: Robust Self-Supervised Stereo Matching with Pseudo-Label Supervision (AAAI 2025 · similarity 0.90)
- Robust Synthetic-to-Real Transfer for Stereo Matching (CVPR 2024 · similarity 0.87)
- Any-Stereo: Arbitrary Scale Disparity Estimation for Iterative Stereo Matching (AAAI 2024 · similarity 0.87)
- Selective-Stereo: Adaptive Frequency Information Selection for Stereo Matching (CVPR 2024 · similarity 0.85)
- Hierarchical Object-Aware Dual-Level Contrastive Learning for Domain Generalized Stereo Matching (NeurIPS 2024 · similarity 0.85)
- All-in-One: Transferring Vision Foundation Models into Stereo Matching (AAAI 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.