A comparative study on point cloud down-sampling strategies for deep learning-based crop organ segmentation

The 3D crop data obtained during cultivation is of great significance to screening excellent varieties in modern breeding and improvement on crop yield. With the rapid development of deep learning, researchers have been making innovations in aspects of both data preparation and deep network design for segmenting plant organs from 3D data. Training of the deep learning network requires the input point cloud to have a fixed scale, which means all point clouds in the batch should have similar scale and contain the same number of points. A good down-sampling strategy can reduce the impact of noise and meanwhile preserve the most important 3D spatial structures. As far as we know, this work is the first comprehensive study of the relationship between multiple down-sampling strategies and the performances of popular networks for plant point clouds. Five down-sampling strategies (including FPS, RS, UVS, VFPS, and 3DEPS) are cross evaluated on five different segmentation networks (including PointNet + + , DGCNN, PlantNet, ASIS, and PSegNet). The overall experimental results show that currently there is no strict golden rule on fixing down-sampling strategy for a specific mainstream crop deep learning network, and the optimal down-sampling strategy may vary on different networks. However, some general experience for choosing an appropriate sampling method for a specific network can still be summarized from the qualitative and quantitative experiments. First, 3DEPS and UVS are easy to generate better results on semantic segmentation networks. Second, the voxel-based down-sampling strategies may be more suitable for complex dual-function networks. Third, at 4096-point resolution, 3DEPS usually has only a small margin compared with the best down-sampling strategy at most cases, which means 3DEPS may be the most stable strategy across all compared. This study not only helps to further improve the accuracy of point cloud deep learning networks for crop organ segmentation, but also gives clue to the alignment of down-sampling strategies and a specific network.