Incremental Instance-Oriented 3D Semantic Mapping via RGB-D Cameras for Unknown Indoor Scene

Table 2

Comparison to the 3D semantic instance segmentation approach from Voxblox++ [16] proposed by Grinvald et al. For 10 sequences from the SceneNN dataset [21], the per-class average precision (AP) is computed using an intersection over union (IoU) threshold of 0.5 over the predicted 3D segmentation masks.

Seq. ID	Method	Bed	Chair	Sofa	Table	Books	Refrigerator	TV	Toilet	Bag
011	Voxblox++	—	75	50	100	—	—	—	—	—
	Ours	—	68.7	67	100	—	—	—	—	—
016	Voxblox++	100	0.0	0.0	—	—	—	—	—	—
	Ours	75	0.0	0.0	—	—	—	—	—	—
030	Voxblox++	—	54.4	100	55.6	14.3	—	—	—	—
	Ours	—	76	100	50	8.3	—	—	—	—
061	Voxblox++	—	—	100	33.3	—	—	—	—	—
	Ours	—		59.9	33.3
078	Voxblox++	—	33.3	—	0.0	47.6	100	—	—	—
	Ours	—	50	—	100	54.2	75
086	Voxblox++	—	80	—	—	0.0	—	—	—	0.0
	Ours	—	66.7	—	—	25	—	—	—	50
096	Voxblox++	0.0	87.5	—	37.5	0.0	—	0.0	—	50
	Ours	0.0	55.7	—	39.5	11.1	—	0.0	—	68.7
206	Voxblox++	—	58.3	100	60	—	—	—	—	100
	Ours	—	60	100	55	—	—	—	—	100
223	Voxblox++	—	12.5	—	75	—	—	—	—	—
	Ours	—	16.7	—	75	—	—	—	—	—
255	Voxblox++	—	—	—	—	—	75	—	—	—
	Ours	—	—	—	—	—	75	—	—	—