Evaluation of the Effect of Transfer Learning to Multi-Instance Detection of Monkeys

Multiple object tracking is an open problem in computer vision. Various studies have presented effective and efficient architectures for object tracking and have paved the way towards the construction and development of task-specific assistive tools such as in animal behavior analysis or road traffic monitoring. Despite the growing interest in computer vision systems for animal behavior analysis, no such tools have been developed for multi-animal detection specific to macaques. In this study, we aim to develop a robust Japanese macaque detection model. We also explore how transfer learning affects the detection accuracy of the trained models. With a mean AP
50
of 83.17%, F1 score of 84%, and a mean IOU
AP@50
of 70.27%, our Japanese macaque detection model pre-trained on the MacaquePose dataset using YOLOv4 yielded the best performance. Transfer learning from a related task increased the mean average precision at IOU
50
by 8% and significantly reduced training convergence time.