Facade Semantic Segmentation Benchmark

ZAHA

We observe that albeit deep-learning-driven methods achieve high performance on various tasks, the facade semantic segmentation still poses an unresolved challenge.

See the ZAHA WACV’25 paper introducing the challenge. Would you like to try and solve it? Please grab the data and report your results!

Facade Semantic Segmentation at LoFG3

LoFG stands for Level of Facade Generalization. See more in the ZAHA paper introducing the concept.

	PointNet	PointNet++	Point Transformer	DGCNN
OA	59.9	66.4	75.0	71.1
P	46.1	37.8	52.7	53.6
R	42.2	35.9	54.7	45.8
F1	38.7	34.8	52.1	44.5
IoU	26.4	25.6	41.6	33.4
Class scores
wall	61.1	68.5	76.8	83.8
window	25.6	26.3	43.1	64.1
door	13.5	7.8	19.8	21.6
balcony	25.1	0.0	77.5	66.7
molding	22.5	43.4	58.0	57.5
deco	0.0	0.0	5.0	0.0
column	22.4	33.4	0.0	37.2
arch	19.2	25.4	50.2	2.6
stairs	16.0	0.0	7.5	5.6
ground surface	12.0	0.0	24.4	21.3
terrain	53.5	53.5	57.6	68.0
roof	18.7	6.8	66.3	57.4
blinds	4.6	2.3	18.5	20.0
interior	59.7	69.1	72.8	88.0
other	42.7	47.1	70.6	74.1

Exemplary results of the deployed semantic segmentation networks for LoFG3

Facade Semantic Segmentation at LoFG2

LoFG stands for Level of Facade Generalization. See more in the ZAHA paper introducing the concept.

	PointNet	PointNet++	Point Transformer	DGCNN
OA	71.9	75.5	78.2	82.6
P	69.6	73.0	75.8	80.0
R	68.1	73.0	76.6	81.8
F1	68.1	72.6	76.1	80.4
IoU	55.8	59.8	63.9	68.5
Class scores
floor	92.3	87.6	90.7	92.1
decoration	26.2	47.1	47.0	70.0
structural	60.9	65.5	67.0	85.2
opening	28.2	27.2	36.0	66.2
other el.	71.2	71.6	78.9	88.8