Segmentation of structural parts of rosebush plants with 3D point-based deep learning methods

Table 11 Segmentation results on 8 real rosebush models for all architectures

		PointNet	PointNet++	DGCNN	PointCNN	ShellNet	RIConv
Flower	III	14.94	72.72	6.96	49.76	47.26	53.42
	S+III	7.54	79.17	51.42	60.55	55.80	52.93
	Gain	− 7.40	+ 6.45	+ 44.46	+ 10.79	+ 8.54	− 0.49
Leaf	III	80.94	95.07	83.57	92.16	89.86	88.22
	S+III	79.48	96.36	91.71	93.40	92.18	90.07
	Gain	− 1.46	+ 1.29	+ 8.14	+ 1.24	+ 2.32	+ 1.85
Stem	III	3.03	76.79	24.34	67.67	52.18	34.53
	S+III	7.06	83.05	60.02	68.81	66.73	51.27
	Gain	+ 4.03	+ 6.26	+ 35.68	+ 1.14	+ 14.55	+ 16.74
MIoU	III	32.97	81.53	38.29	69.86	63.10	58.72
	S+III	31.36	86.19	67.72	74.26	71.57	64.76
	Gain	− 1.61	+ 4.66	+ 29,43	+ 4.40	+ 8.47	+ 6.04

Bold stands for "gain in IoU obtained by incorporating synthetic models" as expressed in the caption of the Figure
The first row for each class corresponds to IoU results of networks trained with three real rosebush models (III). The second row for each class gives the IoU results for the case, where the networks were trained with synthetic models and updated with three real rosebush models (S+III). The third row for each class gives the gain in IoU obtained by incorporating synthetic models. The last three rows of the table corresponds to MIoU over all classes

ISSN: 1746-4811