Automatic horse lameness detection through 2D to 3D reconstruction

Abstract: Lameness is a condition that is difficult to treat in horses when discovered too late and is therefore a common cause for culling. Veterinarians often make a diagnosis based on their subjective experience. In this thesis, we investigate whether neural networks can do lameness detection of horses by using the 3D reconstructed model of the horses. We divide the problem into two parts. The first part is about the 3D model reconstruction of the horse in the videos and then we use neural networks to do lameness detection. We also perform experiments on human videos to test the generalization of our idea, reconstructing the 3D human model in the videos and doing action recognition with neural networks. The two frameworks we use are standard LSTM and LSTM with an attention mechanism. The results of the human experiments show that both networks can separate human actions given the 3D human model sequences and some specific joints are pointed out when doing the two-class action classification. The results of animal experiments preliminarily show that the information of the 3D horse model can be used to perform lameness detection and front-limb lameness is more comfortable for the networks to learn compared to hind-limb lameness.