GhostStereoNet: Stereo Matching from Cheap Operations
Haohua Zhang, Yan Li · 2024
Deep convolutional neural networks show excellent performance in stereo vision tasks. However, most of current network architectures are complex and require high hardware resources. To handle this problem, we propose GhostStereoNet, a light weight end-to-end stereo matching network by applying Ghost modules in GwcNet. We extend the 2D Ghost bottleneck to 3$D$Ghost bottleneck and then replace the convolution layers in GwcNet with 2$D$and 3$D$Ghost bottlenecks. Compared with GwcNet-g, the number of model parameters of our architecture is reduced about 4.3 times. The proposed GhostStereoNet's performance has been demonstrated to remain competitive with the most advanced models, as evidenced by experimental results.