PST: Plant segmentation transformer for 3D point clouds of rapeseed plants at the podding stage

Segmentation of plant point clouds to obtain high-precise morphological traits is essential for phenotyping. Although the fast development deep learning has boosted much research on segmentation clouds, previous studies mainly focus hard voxelization-based or down-sampling-based methods, which are limited segmenting simple organs. complex with a high spatial resolution still remains challenging. In this study, we proposed network transformer (PST) achieve semantic rapeseed plants acquired by handheld laser scanning (HLS) resolution, can characterize tiny siliques as main targeted. PST composed of: (i) dynamic voxel feature encoder (DVFE) aggregate features raw resolution; (ii) dual window sets attention blocks capture contextual information; and (iii) dense propagation module final map. We then integrated an instance head in grouping (PointGroup) developed PST-PointGroup (PG) siliques. The results proved that PST-PG achieved superior performance tasks. For segmentation, mean IoU, Precision, Recall, F1-score, overall accuracy were 93.96%, 97.29%, 96.52%, 96.88%, 97.07%, achieving improvement 7.62, 3.28, 4.8, 4.25, 3.88 percentage points compared second-best state-of-the-art position adaptive convolution (PAConv). reached 89.51%, 89.85%, 88.83% 82.53% mCov, mWCov, mPerc90, mRec90, 2.93, 2.21, 1.99, 5.9 original PointGroup. This study extends phenotyping end-to-end way proves deep-learning-based cloud method great potential resolving traits.