State-of-the-art stereo matching models predominantly predict the disparity of only the left (reference) view. However, applications such as per-eye AR/VR rendering, novel view synthesis, or two-view depth fusion require disparity for both views, and need those two maps to agree. The usual workaround of running the network a second time on the horizontally flipped and swapped pair gives two independent predictions that can be inconsistent, especially on transparent surfaces, in textureless regions, and at occlusion boundaries. We introduce cross-injection (CI), a lightweight extension for iterative (refinement-based) models that can be added to any pretrained left-only iterative matcher. It couples the two passes by injecting the explicit, signed left-right inconsistency as an additional input cue at every step of the refinement stage. We empirically show that CI substantially reduces left-right disagreement on every benchmark (more than halving it on Booster for FoundationStereo), while largely preserving sharp object boundaries. Furthermore, using two state-of-the-art stereo matching models, we show that CI also reduces the single-view disparity error, making it a worthwhile extension even if only left disparity is needed.
Figures & tables
Figure 1: Sample predictions with and without our proposed cross-injection (CI) extension. CI can be added to any iterative stereo matcher (here FoundationStereo [ 56 ] ) and substantially improves left-right consistency between two otherwise independent inference passes (left example); it also reduces per-view failure modes (right example). Here, we show two examples from the Booster dataset [ 39 ] where inconsistent predictions arise mainly on reflective and transparent surfaces. Additional visualizations are shown in Figs. 4 , 5 and 6 .
Figure 2: Overview of the proposed cross-injection architecture. Two refinement passes (the left pass and the flipped-swapped right pass), whose feature encoders and GRU share weights, each iteratively update their own disparity. At every step, each pass receives its own masked left-right inconsistency with respect to the other view (Fig. 3 , Eq. 2 ) as an additional input cue, injected through a zero-initialized encoder. Because the encoder emits zero at initialization, augmenting an already-trained refiner with cross-injection reproduces its original (uncoupled) behavior exactly, and the cross-view correction is learned by fine-tuning.
Figure 3: Illustration of the proposed cross-injection mechanism. The detached partner prediction is horizontally flipped and warped into the frame of the own prediction, using the own disparity as the shift field (Eq. 2 ), and subtracted from the own prediction. The resulting signed inconsistency (“consistency vector”) is passed explicitly to the refiner’s GRU block. We do this with a tensor of shape (B,2,H,W) : the two channels are (i) the signed inconsistency with zeros at non-co-visible pixels, and (ii) the binary co-visibility mask derived from warping. This way, the model can distinguish non-visibility from perfect consistency.
Figure 4: FoundationStereo zero-shot predictions on selected ETH3D [ 42 ] samples with and without our cross-injection method. Even though we are not addressing the ambiguity for transparent objects directly, left/right predictions are more consistent in which visible layer to predict (a). The CI model also successfully corrects major (b) and minor (c) disagreements between the views. Background shrubs in (b) lose some degree of detail.
Figure 5: FoundationStereo zero-shot predictions on selected Booster [ 39 ] (half resolution) samples with and without our cross-injection method. CI corrects errors caused by reflections (a), transparency (b+c) and other random failure modes (d).
Figure 6: FoundationStereo zero-shot predictions on selected KITTI-15 [ 35 ] samples with and without our cross-injection method. (a) shows occasional “smearing” effects on thin vertical structures such as traffic signs or street lamps in front of the sky. Interestingly, this failure mode was already present in FoundationStereo, but our method seems to amplify it. (b) shows an example where both models predict equally sharp edges on road signs and high frequency details (vegetation).
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
FS
S 2 M 2 -L
S 2 M 2 -XL
Optimizer
AdamW, β=(0.9,0.999) , wd =0
Peak LR
10−5
2×10−6
2×10−6
LR schedule
step-drop
cosine
cosine
Optimizer updates
50 k
4 k
4 k
Precision
bf16
fp32
fp32
GPUs ( × bs/GPU)
32×4
32×4
32×2
Appendix
Table S.1 : Fine-tuning configuration for the models in Tab. 2 . Each column applies to both the cross-injection run and its matched no-CI control, which differ only in the cross-injection refiner and the left-right consistency loss. All runs start from the public left-only checkpoint with the cross-injection pathway zero-initialized and fine-tune the full network on 320×736 crops. The S 2 M 2 cosine is defined over exactly the number of updates performed, so the learning rate anneals and the final step is the reported one.
Model
1 pass
2 passes
+ CI
CI overhead
(ms)
(ms)
(ms)
vs. 2 passes
vs. 1 pass
FoundationStereo ( 32 iters)
447
896
1111
+23.9%
+148%
S 2 M 2 -L ( 3 iters)
141
283
298
+5.2%
+111%
S 2 M 2 -XL ( 3 iters)
240
480
498
+3.6%
+107%
Appendix
Table S.2 : What cross-injection costs at inference, and what that is a fraction of. Wall-clock at 480×640 in fp32 (the frozen monocular branch stays fp16 ), batch 1, on one A100-SXM4-80GB; median of 36 CUDA-event-timed runs per configuration, the configurations interleaved round-robin so that GPU clock drift cannot be mistaken for overhead. 1 pass predicts left disparity only. 2 passes is the standard way to obtain both views (two weight-shared, uncoupled passes) and is the baseline cross-injection should be judged against, because any both-view predictor pays it: the second pass costs what the first one does ( +100% ). Cross-injection is then charged per refinement step, which is why the same mechanism costs five to seven times less, relative to the two-pass baseline, on models that iterate three times instead of 32 .
Stage (ms)
1 pass
2 passes
+ CI
Feature extraction (both images)
54.6
111.7
113.4
stem
0.311
0.610
0.613
Cost-volume construction
22.4
48.3
49.3
Cost-volume filtering (3D UNet)
92.5
185.0
184.7
Initial disparity decode
0.125
0.250
0.251
Geometry-encoding pyramid
1.97
3.94
3.92
Appendix
Table S.3 : Inference time breakdown. Per-stage GPU time for FoundationStereo at 32 refinement iterations, under the conditions of Tab. S.2 ; CUDA events are recorded inline and resolved after a single synchronization, so no stage boundary stalls the stream. Columns are the three configurations of Tab. S.2 . Refinement accounts for most of a pass, which is why a per-step cost is material for this model. Per iteration and per view, the cue costs 0.33 ms of warp and 1.60 ms of co-visibility mask, against 7.0 ms for one GRU update: the injected cue is a cheap tensor operation, not a second network, and the mask, not the warp, is what it costs. The sum of stages falls 0.8 – 1.6% short of the independently measured end-to-end time, which bounds the unattributed time.
Model
Latency
Middlebury-H
ETH3D
(ms)
BP-2
LR-BP-1
BP-2
LR-BP-1
FoundationStereo ( K=4 )
469
3.43
10.54
0.50
4.98
FoundationStereo ( K=8 )
514
3.12
10.13
0.44
4.92
FoundationStereo ( K=16 )
635
2.91
10.13
0.44
5.49
FoundationStereo ( K=24 )
780
2.84
10.13
0.43
5.57
FoundationStereo ( K=32 )
882
2.77
9.89
0.43
5.59
Appendix
Table S.4 : Accuracy against latency. Latency is measured at 480×640 as in Tab. S.2 , accuracy at each benchmark’s native resolution. Cross-injection at K=16 is better than the baseline at K=32 on every column shown, at a lower latency : the extra per-step cost is more than repaid by needing fewer steps. On the columns shared with Tab. 2 , the K=32 rows reproduce it; the CI row differs by at most 0.03 because it is evaluated in fp32 here.
Figure S.1 : Cross-injection trained from scratch. A matched pair trained end to end for 100 k steps and differing only in the cross-injection refiner and the consistency loss, shown on four benchmarks at all 20 evaluated steps. Color is the benchmark and line style is the arm ( solid with cross-injection, dashed its matched control), so each solid/dashed pair of one color is a controlled comparison. The left panel reports each benchmark’s own disparity metric (legend), the right panel LR-BP-1 throughout. The dotted line marks the learning-rate drop at 80 k and the shaded band the final 20 k. Right: on every benchmark the solid curve lies below its dashed control at every step. Left: cross-injection starts above its control and closes the gap over the run. Vertical axes are logarithmic. These two runs are comparable only to each other.
Middlebury-H
Booster-H
Region
share %
Δ BP-2
noise
share %
Δ BP-2
noise
all valid pixels
100.0
−0.079
0.033
100.0
+0.065
0.088
textureless
22.4
−0.202
0.202
72.9
+0.137
0.073
depth edges
9.7
−0.045
0.020
0.9
−0.083
0.101
occluded
6.7
+0.562
0.013
1.9
+0.293
0.093
out of view
4.0
−0.170
0.247
5.6
−0.194
0.674
Appendix
Table S.7 : Single-view error change by region. Change in BP-2 (percentage points) for the full method against its matched control (rows (e) and (a) of Tab. 1 ) as the mean over the same six checkpoints; negative is better. Share is the fraction of valid-ground-truth pixels the region covers, exact over the full evaluation set. Noise is the same comparison between control checkpoints, which contain no cross-injection. Bold marks effects at least twice the noise. The rows overlap and do not sum to the all-valid-pixels row.
Pixel group
share %
control
+ CI
Δ
noise
Middlebury-H
consistent ( ∣c∣≤1 px)
96.8
0.254
0.260
+0.006
0.002
inconsistent
3.2
5.259
5.036
−0.223
0.017
Booster-H
consistent ( ∣c∣≤1 px)
94.7
1.319
1.313
−0.006
0.055
inconsistent
5.3
16.845
14.505
−2.341
1.158
Appendix
Table S.8 : Where does cross-injection change the error? Non-occluded valid left-view pixels, split by whether the control’s left and right predictions agree with each other ( ∣c∣≤1 px), for rows (e) and (a) of Tab. 1 , pooled over the six checkpoint pairs. Values are mean absolute disparity error in pixels, for the control and under cross-injection, and their difference Δ ( + CI minus control; negative is better). Noise is the same comparison between control checkpoints, which contain no cross-injection. Because the groups are defined by the control’s predictions, Δ is biased by the grouping; the text tests this.
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.
With the advance of deep neural networks, the quality of disparity maps obtained through stereo matching has steadily improved. However, existing stereo matching methods still struggle to preserve fine-grained geometric details, resulting in blurred edges and over-smoothed predictions in challenging regions. To address these limitations, we propose StereoDiffuer, an iterative diffusion-based stereo matching framework that explicitly models geometric details and progressively refines disparity estimates. The framework incorporates a Saliency Attention Perception (SAP) module to extract salient geometric cues, including object boundaries, thin structures, and sharp edges. Confidence-guided SAP features are combined with the initial disparity estimate to condition an iterative denoising diffusion process, which corrects residual disparity errors and restores geometric details suppressed during cost-volume regularization and upsampling. Experimental results on the Scene Flow and KITTI benchmarks demonstrate the effectiveness of the proposed framework and its competitive performance relative to the compared stereo matching methods.
Bohan Li
Shanghai Jiaotong University, No.800 Dongchuan Road, Shanghai, 200240, China · Ningbo Institute of Digital Twin, Eastern Institute of Technology, No. 568, Tongxin Road, Ningbo, 315000, China
While iterative stereo matching achieves high accuracy, its dependence on Recurrent Neural Networks (RNN) hinders edge deployment, a challenge underexplored in existing researches. We analyze iterative refinement and reveal that disparity updates are spatially sparse and temporally redundant. First, we introduce a progressive iteration pruning strategy that suppresses redundant update steps, effectively collapsing the recursive computation into a near-single-pass inference. Second, we propose a collaborative monocular prior transfer framework that implicitly embeds depth priors without requiring a dedicated monocular encoder, thereby eliminating its associated computational burden. Third, we develop FlashGRU, a hardware-aware RNN operator leveraging structured sparsity and I/O-conscious design, achieving a 7.28× speedup, 76.6% memory peak reduction and 80.9% global memory requests reduction over natvie ConvGRUs under 2K resolution. Our PipStereo enables real-time, high-fidelity stereo matching on edge hardware: it processes 320×640 frames in just 75ms on an NVIDIA Jetson Orin NX (FP16) and 19ms on RTX 4090, matching the accuracy of large iterative based models, and our generalization ability and accuracy far exceeds that of existing real-time methods. Our embedded AI projects will be updated at: https://github.com/XPENG-Aridge-AI.