Gradual Interaction Network for Stereo Matching

Aixin Chong, Hui Yin, Qianqian Du, Yanting Liu, Ming Han

Abstract: Existing stereo matching network models have achieved unprecedented state-of-the-art performance by adopting coarse-to-fine strategy. However, most models ignore the potential of improving stereo matching accuracy by utilizing coarse disparity information feedback to feature extraction. To this end, we propose a Gradual Interaction Network (GINet), which is the first binocular stereo matching learning model to establish a bidirectional interactive bridge between feature extraction and disparity estimation. Firstly, a stereo space grade module (SSGM) is proposed as the disparity evaluation mechanism to obtain the disparity confidence maps at coarse scale and the more precise disparity search space. Secondly, a feature filter module based on disparity feedback (DFFM) is proposed to screen features by using coarse-scale disparity confidence map and coarse-scale probability volume. Finally, we adopt the double branch up-sampling disparity regression strategy (Dup-DR) to the probability volume and disparity search space to obtain an accurate disparity map. Experimental results on the SceneFlow, KITTI2012, KITTI2015 and Middlebury datasets show that the performance of our model is greatly improved compared with previous methods.