LiteMatch: Lightweight Zero-Shot Stereo Matching via Cost Volume Stabilization

Modern learning-based stereo matching often relies on computationally expensive 3D convolutional neural networks (CNNs) or massive pre-trained foundation models to achieve high accuracy and cross-domain generalization. This paper asks whether it is possible to achieve robust, real-time stereo matching by stabilizing the feature representation and cost volume at the source, thereby removing the dependency on heavy regularization and external monocular priors.

The authors introduce LiteMatch, a lightweight framework that employs a two-stage training strategy. First, a dual-encoder architecture—consisting of a Cross-View Correspondence Encoder (CVCE) for global interactions and a High-Frequency Encoder (HFE) for structural detail—extracts geometry-aware features. Second, the authors propose a Cost Volume Consistency Loss (CVC-Loss), which applies voxel-wise binary cross-entropy to softmax-normalized cost distributions. This loss forces the model to produce sharp, unimodal disparity probabilities, effectively stabilizing the cost volume early in training. A lightweight, non-iterative refinement module then produces high-resolution disparity maps, with an optional iterative head available for performance-critical comparisons.

LiteMatch demonstrates that lightweight architectures (ranging from 3.36M to 9.58M parameters) can outperform much larger, state-of-the-art models in zero-shot generalization. On the Scene Flow dataset, LiteMatch achieves superior performance with 39x fewer parameters than leading monocular-prior-based models. Furthermore, the framework sets new benchmarks on the ETH3D dataset and shows exceptional robustness across diverse real-world scenarios, including KITTI and DrivingStereo, while maintaining real-time inference speeds (up to 22 FPS).

This research challenges the prevailing trend of using increasingly massive backbones and heavy post-processing to solve stereo matching. By demonstrating that architectural design and targeted loss functions can stabilize the matching process, the authors provide a viable path for deploying high-performance 3D perception on resource-constrained edge devices, such as those used in autonomous robotics and mobile augmented reality.