Deep learning based approaches from semantic point clouds to semantic BIM models for heritage digital twin

Heritage Science

Table 1 Semantic segmentation results of different methods (%)

	mIoU	OA	Build.	Veg.	Tra.	Ground	Mead.	Wall	Rock	Cars	Others	Bridge	Water
PointNet	33.88	73.28	58.40	79.20	55.50	31.20	25.00	10.20	52.50	7.20	7.70	2.90	42.90
PointNet++	21.24	67.03	37.30	72.00	58.30	15.40	3.00	0.90	23.30	0.00	2.90	0.00	20.50
DGCNN	39.06	69.03	63.68	65.08	77.25	44.41	23.30	19.02	62.01	4.21	2.69	5.42	62.62
RandLA-Net	42.53	76.90	64.84	77.72	87.05	53.96	29.75	37.44	56.64	5.54	8.76	0.91	45.27
KP-FCNN	50.50	84.30	79.49	82.77	92.01	53.87	24.77	31.84	73.50	12.24	13.24	27.42	64.30
Ours	53.03	84.61	81.02	83.25	93.02	59.41	30.60	33.80	72.31	18.97	14.04	25.26	71.62