Research On Macaque Pose Estimation Based On Convolutional Neural Networks

Owing to the biological similarity and homology between macaques and humans,macaque has become one of the most widely-used animal stand-ins to humans in the study of disease research and drug development.During drug trials,behavioral states of macaque monkeys are often observed in order to verify the therapeutic effects of drugs.However,historically,the quantitative measuring of macaque behavior requires experts to analyze manually.Though manual analysis to some extent can meet the requirement of quantitative behavioral analysis of macaque,it is time-consuming and costly,also brings the observer deviation and fatigue error.In this paper,based on the methods of human pose estimation in recent years,convolutional neural networks are introduced to estimate the pose of macaque monkeys,which is able to provide detailed quantitative data for the downstream tasks like behavior analysis and exercise statistics of macaques.Firstly,a complete solution is investigated to tackle the scarcity of training data and the complexity and difficulty of macaque poses.Specifically,the solution includes a novel data augmentation strategy called Random Background Augmentation(RBA)and a novel plug-in module called Mask Guided Attention Module(MGAM).The RBA is proposed to leverage the rich diversity of the general large-scale image datasets to randomly generate diverse backgrounds for the macaques training dataset,which is able to enrich the diversity of the macaque dataset so as to guide our model attend to the foreground macaques.With respect to MGAM,it is proposed to tackle the problems brought by the bushy fur and violent body movements of the macaques.Specifically,the MGAM takes advantage of the macaque contour information and learns the natural constraints contained in the monkey skeleton,which is able to help the model locate and classify the keypoints using the learned prior knowledge of monkey skeleton.Secondly,a novel semi-supervised approach called smoothness-based spatiotemporal consistency learning(SSTCL)and a dual network structure(DNS)are proposed to leverage the amounts of unlabeled real images.Specifically,the SSTCL introduce the smoothness assumption to help the model generalize from the seen training images to the unlabeled images,and the spatio-temporal consistency is designed to leverage both spatial and temporal consistencies to pick the most reliable pseudo labels.And the proposed DNS is proposed to empower the model the ability of self-correction,which can prevent the degeneration caused by the noisy pseudo labels in semi-supervised learning.Finally,the proposed methods are evaluated on the macaque pose dataset.The experimental results show that:(1)the proposed data augmentation method RBA and attention module MGAM can effectively improve the performance of the model.(2)the SSTCL and DNS can achieve excelled performance than the supervised learning,and the promotion is more obvious when there are less labeled images.