关键词: 立体匹配, 域不变特征, 跨域泛化, 特征分布模拟, 代价体构建, 视差上采样

Abstract: To address the issue of cross-domain accuracy degradation in stereo matching methods, a deep stereo matching model, SFENet(Spatial noise Feature Excitation stereo matching Network), is proposed. First, a Hierarchical Visual Transformation (HVT) method is used to discretize the training domain feature distribution and spatial correlation noise is introduced at the  pixel visual transformation layer to simulate the image feature distribution of the real environment, improving the model's adaptability to different target domains. Second, a dual-dimension feature excitation mechanism is designed, enhancing features from both spatial and channel dimensions during cost volume construction, and performing feature-guided excitation aggregation on group-wise correlation volumes to actively filter and enhance effective features, suppress irrelevant information, and retain more effective information for propagation. Both mechanisms jointly guide the model to learn domain-invariant features with stronger generalization capabilities. Finally, an edge-guided disparity upsampling module enhances detail preservation and global consistency, further improving the model's prediction accuracy. Experimental results show that compared to the benchmark model IGEV(Iterative Geometry Encoding Volume), SFENet achieves better performance on multiple cross-domain test datasets. On the KITTI2012 dataset, SFENet achieves a 6.25% reduction in error rate by 3 pixels at the cost of 0.1s inference speed, achieving competitiveness in both accuracy and speed among mainstream stereo matching methods.

Key words: stereo matching, domain-invariant feature, cross-domain generalization, feature distribution simulation, cost volume construction, disparity upsampling