A New Volumetric CNN for 3D Object Classification Based on Joint Multiscale Feature and Subvolume Supervised Learning Approaches

Table 2

Classification results by different methods on ModelNet40 and ModelNet10 datasets.

Network type	Method	Input	Size	Pretrain	Augmentation	ModelNet40 (%)	ModelNet10 (%)
Single volumetric	MS-VDCNN(Ours)	Volumetric	6.2 M	ModelNet40	24	92.93	95.3
	3DShapeNets [15]	Volumetric	38 M	ModelNet40	12	77.32	83.5
	VoxNet [16]	Volumetric	0.92 M	—	12	83	92
	Voxception [56]	Volumetric	—	ModelNet40	24	90	93.28
	OctNet [23]	Volumetric		—		86.5	90.9
	VRN [56]	Volumetric	18 M	ModelNet40	24	91.33	93.61
	Aniprobing [14]	Volumetric	—	—	60	85.6	—
Ensemble volumetric	NormalNet [17]	Volumetric + norm: vector	6.5 M		20	88.8	93.1
	VRN ensemble [56]	Volumetric	90 M	ModelNet40	24	95.54	97.14
	FusionNet [20]	Volumetric + multiview	118 M	ImageNet ModelNet40	60	90.80	93.1
Point cloud	PointNet [50]	Points	0.45 M	—	—	86.2	89.2
	PointNet++ [52]	Points	—	—	—	—	90.7
	3D Capsule [55]	Points	—	—	—	92.7	94.7
	RS-CNN [54]	Points	1.41	—	—	93.6	—
Multiview	MVCNN [11]	Multiview	—	ImageNet	80	90.10	—
	Ma et al. [46]	Multiview	—	ImageNet	12	91.05	95.29
	3D2SeqViews [47]	Multiview	—	ImageNet	12	93.40	94.71