Multi-branch adjustable bottleneck convolution module and end-to-end stereo matching network

Abstract

The invention discloses a multi-branch adjustable bottleneck convolution module (MAB) and an end-to-end stereo matching network, which are used for estimating parallax of left and right images. The channel number and the receptive field of convolution capture information are adjusted by adjusting the scale coefficients of multiple branches in the MAB module and the expansion rate of cavity convolution of each branch, so that the calculated amount, the income of the data access amount and the information amount of the convolution result are balanced and saved. The MAB module can be used as a lightweight feature extraction module and is widely applied to a deep learning network. The lightweight end-to-end stereo matching neural network is constructed on the basis of an MAB module and 3D expansion thereof, compared with a previous stereo matching neural network, the model parameter quantity and the operation frequency are greatly reduced, but the precision reaches the SOTA level through testing on SceneFlow and KITTI data sets. Therefore, the method is easier to deploy to resource-limited systems such as an embedded platform and wearable equipment.

Description

technical field [0001] The present invention proposes a novel Multibranch Adjustable Bottleneck (MAB) convolution module and its 3D expansion, and a lightweight end-to-end stereo matching neural network (MABNet) based on the module for estimating left and right images parallax. The invention belongs to the field of computer vision stereo matching. Background technique [0002] Stereo matching, that is, calculating the disparity of the left and right images obtained by the binocular camera, can further perform depth estimation. It is widely used in various computer vision fields, including autonomous driving, 3D reconstruction, AR / VR, etc. However, the current traditional algorithm based on image processing uses manually set features and functions, which has great limitations, especially for non-textured areas, edge occlusion, and repeated texture areas, which are prone to matching failures. The algorithm based on deep learning, through the study of a large amount of exist...

AI Technical Summary

Problems solved by technology

Although it has shown good performance, it is difficult to deploy to hardware devices with limited resources due to the large amount of parameters and many operations, and the systems that use stereo matching are often used in automobiles, drones, wearable devices, etc.

Images