FFC-Stereo: Stereo matching of binocular endoscopic images by Fast Fourier Convolution
Wei Li, Ruoqi Lian, Huoling Luo, Weikai Qu, Wenjian Liu, Fucang Jia · 2024
Three-dimensional reconstruction of minimally invasive surgical scenes using binocular endoscopic images plays a pivotal role in developing surgical navigation systems. However, there is a significant scarcity of well-labeled endoscopic datasets, and current stereo matching algorithms often fall short in terms of generalization. Consequently, the accuracy of predicting unseen data remains inadequate for practical applications. To address this challenge, we propose a fast Fourier convolution-based stereo matching model, named FFC-Stereo. The FFC-Stereo model includes a downsampling module, a fast Fourier convolution residual module, and an upsampling module. Experimental results indicate that the incorporation of fast Fourier convolution markedly enhances the model’s generalization performance while preserving a straightforward structure. Furthermore, FFC-Stereo demonstrates superior accuracy and faster inference on unseen datasets when compared to state-of-the-art methods. This advancement underscores the potential of FFC-Stereo in improving the efficacy and reliability of surgical navigation systems. The code is available at: https://github.com/Tobyzai/FFC-stereo.