EFSNet: Lightweight stereo matching through hybrid cost volume fusion and multi-scale refinement

Stereo matching is a critical task for 3D perception in robotics and autonomous driving, offering a cost-effective and dense alternative to active sensors such as LiDAR. A key challenge lies in designing architectures that are both highly accurate and efficient enough for real-time deployment on resource-constrained hardware. In response to this challenge, we present EFSNet, a lightweight yet accurate stereo matching architecture that balances performance and efficiency. Building upon the efficient LightStereo framework, our architecture introduces three key contributions. First, we propose a hybrid 2D–3D aggregation module that synergistically combines the local disparity consistency captured by 3D convolutions with the global, all-to-all context modeled by 2D convolutions. Second, a dual-path refinement stage processes the fused features in parallel, with one path specialized for preserving fine details and another for adding robust estimation in low-texture areas. Finally, we introduce LinearLog loss, a training objective that effectively corrects both small and large errors by combining the properties of logarithmic and linear penalties. Our experiments show that EFSNet achieves a compelling trade-off between performance and speed. Our standard model achieves an EPE of 0.64 with an inference time under 20 ms on the SceneFlow dataset. Furthermore, our lightweight EFSNet-lite variant is the fastest among the compared methods, running at 16 ms while surpassing the accuracy of the previous fastest model.