Range-free disparity estimation with self-adaptive dual-matching

Depth estimation from stereo images is an important task in computer vision. Despite of the great contributions that are made in this field, most matching-based methods still face the limitations brought by a pre-set-fixed disparity range. Stereo matching is reconsidered using a specially designed dual-matching method with a cross-attention mechanism, which liberates the algorithm from manually pre-specified disparity ranges and the performance is guaranteed without re-training when the camera rig varies. Moreover, to tackle the mismatches on edges and details, an exquisite module is designed based on left-right consistency, which further refines the estimated disparity map. The efficient multi-scale aggregation is done with both 2D and 3D convolutional layers and the proposed method is proved to be competitive and effective by experiments conducted under popular benchmarks.