EfficientStereo: A Real-Time Stereo Matching Approach Using Lightweight Feature Extraction and Disparity-Dimensional Convolution
Jianfeng Tang, Jingxian Liu, Siran Ding, Yan Pan, Mai Xu · 2025
In this paper, we propose an efficient stereo matching network that achieves competitive performance while significantly reducing computational complexity. Unlike many methods that rely on complex backbone networks, our network employs only a few convolutional layers to extract features, thereby preserving image texture information as much as possible. Subsequently, we construct a 4D cost volume using Group-wise Correlation (GWC) and perform cost aggregation through a 3D convolutional module that incorporates depthwise disparitydimensional convolution. We compare the proposed network with state-of-the-art models on multiple benchmarks, demonstrating its significant advantages in terms of speed, accuracy, and resource utilization. Furthermore, comprehensive runtime analysis demonstrates our method's efficiency in real-time applications. Our model outperforms most real-time models on the KITTI dataset and demonstrates competitive performance on the Scene-Flow dataset. The code has been made publicly available at https://github.com/AbyssFENG/EfficientStereo.