Out-of-distribution- and location-aware PointNets for real-time 3D road user detection without a GPU

Journal of Big Data

Table 6 A comparison of ground segmentation methods

Method	Dataset	Scans	Time (s)	Precision	Recall	Accuracy	IOU
HD [44]	SemKITTI	23201	0.306	0.47	0.95	–	0.45
LF [44]	SemKITTI	23201	0.658	0.38	0.77	–	0.34
GPF [44]	SemKITTI	23201	31.71	0.67	0.63	–	0.45
GPF-Opti [44]	SemKITTI	23201	0.207	0.66	0.59	–	0.43
GPF-RANSAC [44]	SemKITTI	23201	0.028	0.65	0.88	–	0.74
Hybrid-reg [45]	KITTI	5	0.888	–	–	0.88	–
CNN-method [46]	Custom	252	0.139	0.93	0.99	–	–
CRF-method [47]	SemKITTI	3040	0.147	0.80	–	–	0.78
GndNet [48]	SemKITTI	3040	0.018*	0.84	0.99	–	0.84
Ours	SemKITTI	23201	0.016	0.89	0.93	0.93	0.83

Our results are achieved with a single core of a 4.0 GHz Intel i5-7600K 5th generation CPU. 360\(^{\circ }\) scan is divided evenly into 16 sectors. * – GPU inference