Human Detection Under Arbitrary Poses Based On Multiple Instance Learning

Most current research on human body detection focuses only on a few common human body poses with human body in upright positions,while in the real world human bodies may exhibit very rich pose variations(e.g.,when people are bending,sleeping,or sitting).The problem of weakly supervised human body detection under difficult poses(e.g.,multi-view and/or arbitrary poses)is studied.This not only imposes great challenges on the task of human detection,but also makes the job of manual annotation even more difficult and usually only weak annotations are available in practice.The multi-instance learning method relaxes the requirements of accurate labeling and hence being commonly used to address the task.In this paper we study the problem of weakly supervised human detection under arbitrary poses within the framework of multi-instance learning(MIL).Our contributions are four folds:(1)We carry out the implementation and comparison of many MIL algorithms,and analyze and summarize the characteristics and the applicable conditions of all algorithms to provide a reference for the use of MIL algorithms;(2)we identify several crucial factors that may significantly influence the performance,such as the usefulness of a small amount of supervision information,the need of relatively higher RoP(Ratio of Positive Instances),and so on-these factors are shown to benefit the MIL-based weakly supervised detector but are less studied in the previous literature;(3)we first show that in the context of weakly supervised learning,some commonly used bagging tools in MIL such as the Noisy-OR model or the ISR model,tend to suffer from the problem of gradient magnitude reduction when the initial instance-level detector is weak and when there exist large number of negative proposals,resulting in extremely inefficient use of training examples.We hence advocate the use of more robust and simple Max-Pooling rule under such circumstances;(4)we propose a new selective weakly supervised detection(SWSD)algorithm.We also annotate a new large-scale data set called LSP/MPII-MPHB,on which and another popular benchmark dataset we demonstrate the superiority of the proposed method compared to several previous state-of-the-art methods.