- [pdf] [code]
PBCStereo: A Compressed Stereo Network with Pure Binary Convolutional Operations
Although end-to-end stereo matching networks achieve great performance for disparity estimation, most of them require far too many floating-point operations to deploying on resource-constrained devices. To solve this problem, we propose PBCStereo, the first lightweight stereo network using pure binarized convolutional operations. The degradation of feature diversity, which is aggravated by binary deconvolution, is alleviated via our novel upsampling module (IBC). Furthermore, we propose an effective coding method, named BIL, for the insufficient binarization of the input layer. Based on IBC modules and BIL coding, all convolutional operations become binary in our stereo matching pipeline. PBCStereo gets 39x saving in OPs while achieving comparable accuracy on SceneFlow and KITTI datasets.