Seamless synthesis of 3D models from RGB-D data for dairy cow identification

Abstract

In order to monitor health of dairy cows, we previously proposed a system for individual identification from cameras installed on the ceiling of barns, which uses deep learning to identify individual cows with spot patterns as a clue. In this system, to automate the construction of the database of spot patterns, we have developed a method for generating a 3D model from multiple RGB-D data sets. However, the generated images from the 3D model have seams that do not exist in real cows. To solve the problem, we propose a method for synthesizing seamless 3D models. The method defines a distance metric to determine the neighborhood point considering the normal direction. Using the distance, we blend two 3D point clouds by averaging nearest neighbor points and synthesize a seamless point cloud. To confirm the effectiveness of the proposed method, we evaluate the performance of individual identification that uses the images generated by the proposed method.