Tangerine Volume Estimation by Point Cloud Data with Neural Networks

Yield volume estimation is the integral part of modern farming. While classical model-based approaches are already well-developed, in the recent years neural network-based end-to-end methods gain traction. This work is devoted to the case study of applying a PointNet++-like neural network to the problem of tangerine volume estimation with the point cloud data. A set of tangerines filmed by Intel RealSense camera was collected and marked. The dataset includes RGB images and depth images. The volume estimation method is run on this data presented in the form of point clouds. A number of experiments on this dataset were conducted. Experimental results suggest that the problem of volume estimation on the real data can be solved by the means of a single modern computer in real time. The output quality was assessed via comparing the prediction of the method with the real volume of the fruit. The average volume estimation error slightly exceeds 10%.