Pig Datasets of Livestock for Deep Learning to detect Posture using Surveillance Camera

This paper proposes the pig Datasets for deep learning to detect a posture using a surveillance camera. The proposal aims to develop the best quality Datasets that can be input to the object detection pipeline for building a deep learning model. The proposed Datasets have two types with 7 and 9 categories(each called Class 7 and Class 9), and each evaluated. In Class 9, one of the categories labels to the ground-truth bounding box as a fake data assuming that the hidden head of a pig. The hidden head of pig shows to decrease performance. When training YOLOv2 and SSD against proposed Datasets, the performance an average precision (AP) of each training result evaluates the trained model with three types of optimization algorithms: Adam, SGDM, and RMSProp. The detection accuracy for proposed Datasets Class 7 reaches 97%. YOLOv2 using the proposed Datasets of Class 7 shows seven times better performance than SSD.