Better feature network for Iterative Geometry Encoding Volume

Iterative Geometric Encoding Volume for Stereo Matching (IGEV-Stereo) has shown excellent performance in stereo matching tasks. It is proposed on the basis of Recurrent All Pair Field Transform of Optical Flow (RAFT) and exhibits strong inference effects. Although IGEV-Stereo solves the problem of lack of non local geometric information and difficulty in handling local fuzziness in all pairs correlations, in the feature extraction stage, the cost of constructing the feature network and contextual information network proposed by IGEV-Stereo still lacks inference details and fuzzy area information. In this article, we use a novel feature extractor structure that simultaneously implements both feature network and context network functions, which is a new stereo matching feature extractor structure. This comprehensive feature extractor, based on the IGEV structure, provides stronger performance for the algorithm in processing stereo encoding geometry and contextual information, as well as local matching details. The newly used structural model achieved good results in Scene Flow and KITTI 2012 and 2015, and achieved better results compared to the previous IGEV-Stereo model.