RetinaStereo: Dynamic-Volume Stereo Matching Network

Abstract

Existing stereo matching techniques often struggle with detailing subtle objects on depth edges. To alleviate this problem, we introduced the Dynamic-Range Disparity Initialization module, which integrates three complementary branches: the dynamic dense volume for localized disparity sampling, the sparse global volume for encoding search center information, and the background static volume with skip connections for enhancing depth edge accuracy. The dynamic dense volume identifies optimal search centers for each pixel and performs local neighborhood sampling for cost aggregation, thereby generating the initial disparity map. Since the search center positions get lost while assembling the disparity samples, a sparse global volume is proposed to implicitly encode these positions during the training process of the network. We also designed the Inception Update module based on our analysis of convolutional structures and nonlinear gating mechanisms. RetinaStereo achieves state-of-the-art performance on the KITTI-2015 leaderboard for the D1-fg metric among published methods.