HCVNet+: A Lightweight and Accurate Network for Binocular Stereo Matching

Binocular stereo matching, the foundation of many visual and robotic applications, typically estimates disparity and depth from the cost volume constructed by the left and right feature maps. Although most learning-based models can achieve superb effects, it is still difficult to satisfy the demand for a better trade-off between accuracy and efficiency. Thus, our main aim is to enable the backbone model become lightweight and efficient without sacrificing accuracy. Firstly, to achieve saving in computation, we present the 3D effective channel attention module(3DECAM) to make the target cost volume fetch more helpful information from the source cost volume. Then, in order not to greatly affect the accuracy, we design the residual-block-style attention hourglass module(RAHM) to appropriately aggregate and utilize the input cost volume. At last, combining these ideas, we construct a lightweight and efficient end-to-end stereo matching network called HCVNet+. Under the same running condition, HCVNet+ can achieve 0.669 EPE with 220ms avg run time and 135G MACs, which is faster than HCVNet(260ms avg run time), also descending HCVNet’s MACs(268G) by almost 50%.