Re-parameterization Making GC-Net-style 3DConvNets More Efficient
For depth estimation using a stereo pair, deep learning methods using 3D convolution have been proposed. While the estimation accuracy is high, 3D convolutions on cost volumes are computationally expensive. Hence, we propose a method to reduce the computational cost of 3D convolution-based disparity networks. We apply kernel re-parameterization, which is used for constructing efficient backbones, to disparity estimation. We convert learned parameters, and these values are used for inference to reduce the computational cost of filtering cost volumes. Experimental results on the KITTI 2015 dataset show that our method can reduce the computational cost by 31-61% from those of trained models without any performance loss. Our method can be used for any disparity network that uses 3D convolution for cost volume filtering.