Abstract

This tutorial first introduces higher-order statistical modeling in shallow architectures and its prelimary combination with deep architectures. Then we focus on state of the art on higher-order statistical modeling in deep Convolutional Netural Networks (ConvNets). The higher-order statistical modeling based deep networks signfincantly improve performance of mainstream, first-order ConvNets in a variety of vision taksks, including large-scale visual recogniton, small-scale finegrained visual categorization, and object detection, among others. The effectiveness of of higher-order modeling provides motivating perspective and idea to develop novel model, architecture and theory of deep ConvNets.

Contents

Overview: What is and Why Study Higher-order Statistical Modeling?
Presented by Peihua Li, slides

Part I Classical Higher-order Statistical Modeling and Its Prelimary Combination with ConvNets
Presented by Qilong Wang, slides

Part II Global Second-order Pooling and Distribution Pooling Deep ConvNets
Presented by: Peihua Li, Qilong Wang, slides

Part III Approximate Higher-order Pooling Networks
Presented by: Wangmeng Zuo, slides

Part IV Code and Achievements in Alibaba Large-scale Image Search Challenge and Large-scale iNat Challenge@FGCV5 CVPR18
Presented by: Qilong Wang, Jiangtao Xie, Peihua Li, slides