Robust Stereo Matching Using Noise Feature Extraction Network with Dynamic Convolution

Stereo matching in noisy scenes presents significant challenges, as interference noise blurs the image, leading to ambiguities in matching. Most existing methods attempt to denoise the images prior to performing stereo matching. However, these approaches are not ideal for stereo matching, as denoising algorithms may not only eliminate noise but also valuable texture information, consequently causing matching errors. Furthermore, previous methods struggle to utilize depth cues from the noise effectively, leading to inefficient cost aggregation during stereo matching. To address these issues, we introduce a noise feature extraction network with dynamic convolution (NFENDC) that can extract depth cues from noisy images and improve stereo matching using these cues. In addition, we devise a noise volume representation to systematically capture these depth cues from the noise and fuse it with the cost volume for improved cost aggregation, rectifying the ambiguous matching caused by noise. Experimental results on KITTI benchmarks substantiate that our proposed method significantly enhances the matching performance of state-of-the-art algorithms in noisy scenes.