Design and implementation of an enhanced interval geometric encoding volume-based stereo matching network

This paper addresses the problem of ambiguous disparity prediction in edge and low-texture regions of the Interval Geometric Encoding Volume network and proposes a method that integrates multi-disparity interval geometric encoding with the Squeeze-and-Excitation attention mechanism. The proposed method constructs a geometric encoding volume through multi-disparity interval grouping to adaptively handle disparity requirements across different ranges. A lightweight Squeeze-and-Excitation module is incorporated to dynamically suppress irrelevant features and strengthen key geometric semantic information. In addition, a Convolutional Gated Recurrent Unit is introduced to iteratively update the disparity field, and the Hourglass module is used to optimize disparity discontinuity. Experiments are conducted on the Scene Flow dataset to analyze the effects of the proposed components on disparity estimation performance and computational efficiency.