Coarse-to-fine stereo matching enables efficient disparity estimation but remains sensitive to errors in propagated disparity, which may place ground-truth correspondences outside subsequent local search ranges. We propose TRUST-Stereo (Teacher-Assisted Robust Uncertainty-Guided Search and Transfer for Fast Stereo Matching), which addresses this limitation through an Uncertainty-Guided Residual Search Module (UGRSM). UGRSM estimates the reliability of propagated disparity and predicts residual offsets to adaptively relocate local cost-volume centers. Additionally, we introduce a two-stage Senior-Junior Teacher Assistant knowledge distillation strategy to transfer rich monocular representations from a large foundation model to a lightweight backbone to achieve fast stereo matching performance. Experiments on KITTI 2012, KITTI 2015, ETH3D, and Middlebury demonstrate strong zero-shot generalization. Using only 1.4M synthetic stereo pairs, TRUST-Stereo achieves state-of-the-art results on several benchmark metrics, outperforming competing efficient methods trained with substantially larger datasets, including real-world pseudo-labeled samples. Specifically, it achieves a KITTI 2015 EPE of 0.93, ETH3D Bad 1.0 of 2.13% and Middlebury Bad 2.0 of 6.79% while maintaining inference at approximately 50 ms per stereo pair.
Figures & tables
Figure 1: Zero-shot generalization comparison under SceneFlow-only training.
Figure 2: Overview of TRUST-Stereo. A coarse-to-fine stereo matching architecture iteratively refines disparity using ConvGRUs. The proposed UGRSM adaptively adjusts local search centers, while Senior–Junior Teacher Assistant distillation provides lightweight monocular guidance for robust zero-shot generalization.
Figure 3: Overview of the proposed two-stage Senior–Junior Teacher Assistant knowledge distillation framework.
Figure 4: Architecture of the proposed Uncertainty-Guided Residual Search Module (UGRSM). The module predicts an uncertainty map and a residual search offset from the propagated disparity and image features. The uncertainty score modulates the residual offset to produce an adjusted local cost-volume center used for local disparity search.
Method
KITTI 2012
KITTI 2015
ETH3D
Middlebury
Runtime (ms)
D1
EPE
D1
EPE
Bad 1.0
Bad 2.0
Efficient methods: SceneFlow training
CoEX [ 1 ]
22.30
3.37
17.33
3.00
31.97
26.42
27
MobileStereoNet-2D [ 36 ]
19.30
2.51
21.88
2.85
17.10
37.98
41
MobileStereoNet-3D [ 36 ]
19.59
2.67
17.89
3.01
18.42
25.26
67
FastACV+ [ 51 ]
16.69
2.29
15.28
3.27
11.50
27.34
45
Table 1: Zero-shot generalization on KITTI 2012, KITTI 2015, ETH3D, and Middlebury-H. Baseline results are adopted from [ 22 ] , except Fast Foundation Stereo R , which we retrained on SceneFlow since the official SceneFlow-only weights are not released. ETH3D and Middlebury-H results are evaluated on valid non-occluded pixels. SceneFlow-only methods are trained on SceneFlow. In million-scale training, ∗ denotes models retrained using the synthetic data and training strategy of [ 18 ] ; † denotes 30M pseudo-labeled samples; ‡ denotes 1.8M synthetic and 0.5M real-world pseudo-labeled samples; and \lx@sectionsign denotes our 1.4M synthetic-only training. Best and second-best results within each setting are highlighted in red (bold) and yellow, respectively.
Method
KITTI 2012
KITTI 2015
MACs (G)
3-noc
3-all
4-noc
4-all
EPE-noc / all
D1-bg
D1-fg
D1-all
DispNetC [ 28 ]
4.11
4.65
2.77
3.20
0.9 / 1.0
4.32
4.41
4.34
–
DeepPruner-Fast [ 11 ]
–
–
–
–
–
2.32
3.91
2.59
194
AANet+ [ 52 ]
1.55
2.04
1.20
1.58
0.4 / 0.5
1.65
3.96
2.03
–
DecNet [ 55 ]
–
–
–
–
–
2.07
3.87
2.37
–
BGNet+ [ 48 ]
1.62
2.03
1.16
1.48
0.5 / 0.6
1.81
4.09
2.19
77
Table 2: Fine-tuned comparison on the KITTI 2012 and KITTI 2015 test sets. Lower values indicate better performance. Baseline MACs are adopted from [ 22 ] and reported at 1242×375 input resolution.
Figure 5: Qualitative zero-shot disparity estimation comparison on ETH3D and Middlebury-H. The first two rows show ETH3D, while the last two rows show Middlebury-H. Disparity predictions from TRUST-Stereo (ours), Fast-FoundationStereo, LightStereo-M, and BANet-2D are compared. All models are trained exclusively on SceneFlow and evaluated zero-shot without target-domain fine-tuning.
Dataset
w/o UGRSM (%)
w/ UGRSM (%)
KITTI 2012
0.2833
0.2701
KITTI 2015
0.5189
0.3188
ETH3D
2.7000
1.4300
Middlebury
4.0777
2.6741
Table 3: Effect of UGRSM on local cost-volume coverage under SceneFlow-only zero-shot generalization. Values indicate the percentage of ground-truth correspondences outside the local search range. Lower is better.
Dataset
Metric
w/o UGRSM
UGRSM (No Uncertainty)
Full UGRSM (Ours)
KITTI 2012
D1 (%)
4.81
4.85
4.64
KITTI 2015
D1 (%)
6.79
6.73
6.23
ETH3D
Bad 1.0 (%)
9.62
13.46
9.12
Middlebury
Bad 2.0 (%)
15.96
14.82
14.62
Table 4: Effect of uncertainty-guided residual search on zero-shot disparity accuracy. All variants are trained on SceneFlow without knowledge distillation. ETH3D and Middlebury are evaluated on all valid pixels, including occluded regions. Lower is better.
Dataset
Metric
Direct KD
ViTKD
SJTA KD (Ours)
KITTI 2012
EPE
0.95
0.99
0.93
D1 (%)
4.45
4.96
4.30
KITTI 2015
EPE
1.19
1.18
1.16
D1 (%)
5.71
6.35
5.78
ETH3D
Bad 1.0 (%)
24.80
21.31
8.79
Bad 2.0 (%)
20.43
16.55
4.35
Table 5: Zero-shot comparison of direct Teacher–Student KD, ViTKD, and the proposed Senior–Junior TA KD. Lower is better.
SJTA
UGRSM
KITTI 2012
KITTI 2015
D1 (%) ↓
D1 (%) ↓
×
×
4.81
6.79
×
✓
4.64
6.23
✓
×
4.60
6.43
✓
✓
4.30
5.78
Table 6: Ablation study of the proposed Senior–Junior Teacher Assistant (SJTA) distillation and Uncertainty-Guided Residual Search Module (UGRSM) under SceneFlow-only training. All variants are evaluated zero-shot on KITTI 2012 and KITTI 2015 using D1 (%). Lower is better.
Real-time stereo matching is crucial for robotics, autonomous systems, and embedded vision applications, where both computational efficiency and disparity accuracy are required. Recent coarse-to-fine stereo matching methods improve efficiency by progressively refining disparity estimates using local cost volumes at higher resolutions. However, these methods rely heavily on the accuracy of propagated disparity estimates from previous stages. When the propagated disparity is inaccurate, the ground-truth correspondence may fall outside the predefined local search range, leading to unrecoverable matching failures during subsequent refinement. In this paper, we propose URS-Stereo, a real-time coarse-to-fine stereo matching framework that addresses this limitation through uncertainty-guided search adaptation. Specifically, we introduce an Uncertainty-Guided Residual Search Module (UGRSM), which predicts the reliability of propagated disparities together with residual search offsets to adaptively relocate the centers of local cost volumes before disparity refinement. By dynamically adjusting the search region according to the confidence of the propagated disparity, the proposed method significantly improves the robustness of local correspondence estimation while preserving the computational efficiency of coarse-to-fine stereo matching. Extensive experiments on SceneFlow, KITTI 2012, KITTI 2015, Middlebury, and ETH3D demonstrate that URS-Stereo consistently improves disparity estimation while maintaining real-time inference speed, validating the effectiveness of the proposed uncertainty-guided search strategy
Pouya Sohrabipour, Chaitanya kumar reddy Pallerla, Dongyi Wang
Existing iterative stereo matching methods primarily adopt two types of correspondence representation: explicit matching search via correlation volumes and local residual refinement via warped features, yet the two remain separately modeled. We propose WAVE-Stereo, built on a core insight: correlation volumes and feature warping provide complementary matching cues. GeoWarp Correspondence Encoder (GWCE) encodes matching search, residual alignment, and disparity prior in parallel at the ConvGRU input. To mitigate matching degradation in textureless regions, we propose Periodic Global Context Propagation (PGCP), which propagates global spatial information in a periodic manner. On five real-world benchmarks -- Middlebury, ETH3D, KITTI 2012, KITTI 2015, and Booster -- WAVE-Stereo achieves competitive zero-shot generalization accuracy without any external foundation model prior, achieving 3.18% D1-all on KITTI 2015, 4.42% Bad-2.0 on Booster, and 66ms real-time inference, striking a favorable balance between accuracy and efficiency. Our code is available at https://github.com/yamanoko-do/WAVE-Stereo.
Zehan Liu, Yage He, Xianwu Gong
School of Electronics and Control Engineering Chang’an University Xi’an, China
Standard attention mechanisms are not well suited to stereo matching. Global attention scales quadratically and provides no explicit matching constraint, while local attention is efficient but loses long-range correspondences. We propose MatchAttention, an attention mechanism that embeds an explicit matching constraint into attention by treating the relative position between a query and its matched key as a learnable component of attention sampling. Centering a small contiguous sampling window on this learnable relative position enforces the matching constraint and supports long-range correspondence at strictly linear attention complexity. A differentiable contiguous attention sampling (CAS) operator enables sub-pixel accuracy, and cascaded MatchAttention blocks iteratively refine the relative positions through residual connections. We instantiate MatchAttention as a hierarchical coarse-to-fine stereo network with two variants. MatchAttentionXL targets accuracy and MatchAttentionRT targets real-time edge inference. MatchAttentionXL achieves state-of-the-art accuracy on Middlebury V3 and top results across KITTI 2012/2015 and ETH3D. MatchAttentionRT runs at 9.3 ms on RTX 4060 Ti and 79.1 ms on Jetson Orin NX 16 GB at 1024 x 512, making it the first stereo model to deliver real-time edge inference without sacrificing zero-shot generalization. The code is available at https://github.com/TingmanYan/MatchAttention.
Tingman Yan, Tao Liu, Chenghao Li +4
School of Control Science and Engineering, Dalian University of Technology, Dalian, 116024, China · Pika Labs · Department of Automation, Shanghai Jiao Tong University, Shanghai, 200240, China +1