While usage of convolutional neural networks (CNN) is widely prevalent, methods proposed so far always have considered homogeneous kernels for this task. In this paper, we propose a new type of convolution operation using heterogeneous kernels. The proposed Heterogeneous Kernel-Based Convolution (HetConv) reduces the computation (FLOPs) and the number of parameters as compared to standard convolution operation while it maintains representational efficiency. To show the effectiveness of our proposed convolution, we present extensive experimental results on the standard CNN architectures such as VGG, ResNet, Faster-RCNN, MobileNet, and SSD. We observe that after replacing the standard convolutional filters in these architectures with our proposed HetConv filters, we achieve 1.5 × to 8 × FLOPs based improvement in speed while it maintains (sometimes improves) the accuracy. We also compare our proposed convolution with group/depth wise convolution and show that it achieves more FLOPs reduction with significantly higher accuracy. Moreover, we demonstrate the efficacy of HetConv based CNN by showing that it also generalizes on object detection and is not constrained to image classification tasks. We also empirically show that the proposed HetConv convolution is more robust towards the over-fitting problem as compared to standard convolution.
Recommended citation: Pravendra Singh, Vinay Kumar Verma, Piyush Rai and Vinay P. Namboodiri, “HetConv: Beyond Homogeneous Convolution Kernels for Deep CNNs”, International Journal of Computer Vision, accepted