Omni-supervised joint detection and pose estimation for wild animals.

• We work on multiple wild animals detection and pose estimation. • We constructed a large wild animal surveillance dataset. • We proposed a ratio based data distillation framework. Monitoring wildlife populations and activities have significance for biology and ecology. With the rapid development of computer vision and deep learning techniques, it is possible to employ state-of-the-art convoluntional neural network (CNN) based detectors to process the big data from field surveillance cameras and assist in the following studies. However, data labelling during the training stage is a very time-consuming, labour intensive and expensive task. In this paper, we detect multiple animals (Kangaroo, emu, dingo, bird and wildcat) in the wild in an Omni-supervised learning setting. The unlabeled data from the surveillance cameras will be filtered and used for training via data distillation approach. Moreover, we also perform joint pose estimation and detection for Kangaroo which has the most samples in the dataset. To study the feasibility, we also built a large full high definition (HD) wild animal surveillance image dataset from collected data from several national parks across the State of Queensland in Australia and this dataset will be made publicly available. Extensive experiments show that the detection and pose estimation results can be further improved by using unlabeled data wisely. [ABSTRACT FROM AUTHOR]