Visual and Visual-Inertial SLAM: State of the Art, Classification, and Experimental Benchmarking

Table 6

Absolute pose error for each tested method on each dataset, averaged on five runs.

APE in cm	Vins-Mono			ORB-SLAM2			DSO			ROVIO			LSD-SLAM
	()			()			()			()			()
	MH 01	8.46 (11.50%)	3.79	20.3	4.3 (1.29%)	2.00	7.93	7.57 (2.47%)	3.54	16.28	30.4 (0.90%)	15.08	96.96	11.76 (12.45%)	8.27	89.92
MH 03	9.51 (7.45%)	4.71	25.25	3.89 (2.68%)	1.71	9.47	10.05 (7.70%)	4.71	23.90	39.25 (0.00%)	16.44	78.03	—	—	—
MH 05	17.39 (2.71%)	7.51	32.01	5.31 (3.21%)	2.19	11.89	13.87 (4.21%)	5.51	24.30	105.45 (0.09%)	49.41	223.36	101 (12.1%)	59.3	722
IRSTV	695	367	2395	649	396	1657	1116	897	5188

Results in italics indicate that the algorithm failed on some of the five runs by losing track. The numbers are thus average on the other runs. The numbers in parentheses are the standard deviation of the RMSE (which is averaged on five runs). Results on IRSTV’s dataset are further explained in the result analysis of each concerned method. None of the methods reconstructed the full IRSTV path but only some parts of it: DSO 292.75 m, ORB-SLAM2 596.80 m, and Vins-Mono 212.83 m. APE was obtained with evo package github.com/MichaelGrupp/evo.