LiteMatch | ECCV 2026

Architecture

LiteMatch features a two-stream encoder with cross-view correspondence and a disparity refinement module. The framework separates high-frequency and semantic features for robust matching, then stabilizes the cost volume through our novel Cross-View Correspondence (CVC) Loss.

3.36M

Total Parameters

2.76

EPE on DrivingStereo

0.05%

Bad1 on Middlebury

Full Network Architecture

Feature Extraction & Stabilization: Dual-stream encoder (E1 for cross-view correspondence, E2 for high-frequency features via FFT filter Bh). The cost volume is refined iteratively using CVC-Loss to produce stable disparity D2.

Disparity Refinement Module: A non-iterative transformer decoder (Tde) with disparity context encoder (Egeo) and correspondence encoder (Ecorr) produces the base disparity DBase, followed by iterative refinement (IR) to generate the final disparity Dfinal.

Visual Results

Qualitative comparisons demonstrating LiteMatch's superior zero-shot generalization across early-stage disparity, parameter efficiency, and fine-grained detail preservation.

Zero-Shot Generalization

(a) Early stage disparity without cost volume post-processing — LiteMatch produces significantly cleaner initial estimates compared to Selective-IGEV and MonSter.

(b) Model size vs. EPE on DrivingStereo — LiteMatch (3.36M) achieves the lowest error while being ~100x smaller than competing methods.

(c) Zero-shot visual results on unseen scenes — our method closely matches ground-truth disparity with sharp object boundaries and accurate depth discontinuities.

Quantitative Results

LiteMatch achieves competitive performance with only 3.36M parameters, outperforming methods with 100x more parameters on zero-shot benchmarks.

Method	Params	DrivingStereo (EPE)	Middlebury (Bad1%)	ETH3D (Bad1%)
MonSter	388M	2.89	0.09	0.55
FoundationStereo	375M	3.12	0.37	0.24
SMoEStereo	113.6M	3.45	0.28	0.30
DEFOM-Stereo	43.2M	4.02	0.28	0.55
Selective-IGEV	13.14M	5.23	-	-
LiteMatch (Ours)	3.36M	2.76	0.05	0.21

Detailed Visual Comparisons

Top: Middlebury benchmark comparisons (Bad1% scores shown). Bottom: KITTI and zero-shot generalization scenarios. LiteMatch produces cleaner boundaries, fewer artifacts, and more accurate disparity maps across all settings.

Citation

If you find this work useful, please consider citing:

@inproceedings{litematch2026,
  title={LiteMatch: Lightweight Zero-Shot Stereo Matching via Cost Volume Stabilization},
  author={Khan, Md R. and Others},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2026}
}