Efficient Fusion of Sparse and Complementary Convolutions for Object Recognition and Detection
We propose a new method for exploiting sparsity in convolutional kernels to create compact and computationally efficient convolutional neural networks (CNNs). Our approach is based on hand-crafted sparse kernels that are spatially complementary, allowing for an effective combination of them to represent more complex but discriminative kernels in an efficient way. Based on this, we develop a module that can be used to replace a convolutional kernel of any size greater than 1. We integrate such a module into various existing CNN models and conduct extensive experiments to demonstrate the effectiveness of our proposed approach on image classification as well as object localization and detection. The experiments validate the adaptability of the proposed method. For classification and localization, the proposed approach achieves competitive or better performance than the baselines and related works for various networks while providing lower computational costs and fewer parameters (on average, a 2-3× reduction of convolutional parameters and a 2-4× speedup in computation). On the other hand, our approach leads to a VGG16-based Faster RCNN detector that is 12.4× smaller and about 3× faster than the baseline.
READ FULL TEXT