The KITTI Vision Benchmark Suite

Stereo image pairs encode 3D scene information through the correspondences between the left and right images. Traditional CNN-based methods typically rely on cost volume techniques to capture stereo correspondences over large disparities but suffer from large parameter sizes and slow computation speeds. Although existing research has optimized the modeling of stereo correspondences, the overall network architecture still needs improvement. To address these challenges, inspired by the significant achievements of Kolmogorov-Arnold Networks (KANs) in terms of accuracy and interpretability, we propose a new unsupervised stereo matching network, WT-KAN. This network reshapes the traditional U-Net architecture by introducing KAN layers, combining nonlinear learnable activation functions, wavelet transforms, and axial attention mechanisms. It enhances the network's ability to capture details and shapes and establishes long-range dependencies. WT-KAN can more efficiently learn complex stereo c

0.2

N/A