A study on 3D reconstruction of animals in the wild

In the current age of technology, 3D reconstruction has become an essential asset in many industries. 3D reconstruction of the human body is used in multiple areas, for instance, in the medical industries for performing training and surgeries. Predicting 3D models from a 2D image is a big challenge as there is limited and insubstantial work available in the 3D reconstruction of animals. It is challenging to capture 3D scans or even 2D images of animals in various motions, shapes and structures in the wild. For the purpose of this thesis, we have performed five experiments and nine testing of various approaches to design a feasible pipeline to overcome the limitation of 3D reconstruction of animals in the wild. We have studied and tested two classes of methods. The first one statistical models of shapes, e.g., SMPL, where the class of shapes being reconstructed is known in advance and represented using a template (mean) and its modes of variation. The second class of methods are generic and are based on depth reconstruction. Furthermore, the study highlights, the limitation of using both the pathways through experiments and results obtained from training the networks using a custom dataset of cows. Index Terms— CNN, SMPL, 3D Human Body, 3D Template Matching, Depth Estimation, Deep Learning