Building3D

Building3D Roof Reconstruction Task

In this evaluation, we utilize precision and recall to assess the performance of each method. To ensure fairness, the evaluation environment remains consistent for every method. Among them, ACO refers to the average corner offset, CP and EP represent the corner and edge precision, CR and ER represent the corner and edge recall, and C-F1 and E-F1 represent the corner and edge F1 scores.

Tallinn City Benchmark

ID	Methods	ACO	CP	CR	C-F1	EP	ER	E-F1
1	Our supervised	0.29	0.9	0.53	0.66	0.88	0.23	0.36
2	Our-self-supervised-80%	0.28	0.87	0.55	0.67	0.89	0.16	0.27
3	Our-self-supervised-50%	0.3	0.84	0.53	0.65	0.85	0.15	0.26
4	Our-self-supervised-20%	0.37	0.76	0.49	0.61	0.78	0.12	0.21
5	Point2RooF	0.39	0.65	0.3	0.41	0.66	0.08	0.14
6	Our-self-supervised-10%	0.39	0.71	0.46	0.56	0.6	0.01	0.02
7	Our-self-supervised-1%	0.57	0.34	0.04	0.07	0.13	0	0

Entry-level Benchmark

ID	Methods	ACO	CP	CR	C-F1	EP	ER	E-F1
1	Our supervised	0.26	0.89	0.66	0.76	0.91	0.46	0.61
2	PointNet++-FeatureExtractor	0.34	0.79	0.52	0.63	0.84	0.33	0.47
3	Point-M2AE-FeatureExtractor	0.24	0.88	0.69	0.77	0.9	0.31	0.46
4	3D-OAE-FeatureExtractor	0.27	0.86	0.68	0.76	0.79	0.32	0.46
5	PAConv-FeatureExtractor	0.33	0.75	0.57	0.65	0.85	0.31	0.45
6	DGCNN-FeatureExtractor	0.32	0.73	0.58	0.65	0.81	0.3	0.44
7	Point-BERT-FeatureExtractor	0.25	0.88	0.69	0.77	0.9	0.29	0.44
8	Our-self-supervised-FeatureExtractor	0.25	0.87	0.69	0.77	0.87	0.27	0.41
9	PointNet-FeatureExtractor	0.36	0.71	0.5	0.59	0.81	0.26	0.39
10	Point2RooF	0.3	0.66	0.48	0.56	0.71	0.26	0.38
11	Point-MAE-FeatureExtractor	0.27	0.85	0.69	0.76	0.86	0.22	0.35
12	Stratified-Transformer-FeatureExtractor	0.38	0.72	0.51	0.62	0.75	0.22	0.34
13	RandLA-Net-FeatureExtractor	0.35	0.7	0.6	0.65	0.67	0.16	0.25