Detection of Animal Behind Cages Using Convolutional Neural Network

Abstract

There are many attempts on detecting animal using Convolutional Neural Network. However, many of them failed to detect animals behind cage bars as the mesh patterns of such bars usually affected a detectability of a detection model. A main hypothesis is that most of existing models trained for detecting animals does not have enough pictures of animals behind cage bars as in a training set. In this paper, panda and deer are used as case examples. The training data is gathered specifically for this research work. The M2Det is used as the main network together with the transfer learning approach and its pretrained weights. In our experiments, it is found that a number of training images of animals behind cage bars greatly affects the detection performance. Also, adding more training images of animals without cages could also improve the performance of the detection model on the same task.