EFFICIENT DEEP CONVOLUTIONAL ARCHITECTURES FOR IMAGE AND VIDEO REPRESENTATION LEARNING

Abstract

This thesis focuses on improving the quality of the learned deep feature representations by designing more efficient network architectures. It starts from improving the micro building blocks of existing deep convolutional neural networks, follows by improving the global topologies of deep convolutional neural networks, and ends with introducing non-convolutional features to compensate the limitations of convolutional networks. The thesis includes Mulit-fiber Unit, Octave Convolution Operator, Dual Path Networks, Group Orthogonal Networks, Collective Learning Unit, Double Attention Unit, and Graph-based Global Reasoning Unit, all of which are proposed during my Ph.D. period. These proposed methods are extensively evaluated on standard benchmark datasets include ImageNet, MS-COCO, Pascal VOC, Cityscapes for 2D image recognition tasks, and Kinetics for 3D video recognition tasks.