Efficient ConvNet Feature Extraction with Multiple RoI Pooling for Landmark-Based Visual Localization of Autonomous Vehicles

Table 2

Localization accuracy comparison of FastRCNN-AlexNet versus AlexNet and FastRCNN-VGG-M1024 versus VGG-M1024 in terms of maximum precision at 100% recall (Pr. at 100% Re.), maximum recall at 100% precision (Re. at 100% Pr.), and average precision (AP). The highest value with respect to each metric on each dataset is highlighted in bold. The middle italic values are the difference between FastRCNN-AlexNet/VGG-M1024 and AlexNet/VGG-M1024.

Method	UACampus			St. Lucia			Nordland			Mapillary
	Pr. at 100% Re.	Re. at 100% Pr.	AP	Pr. at 100% Re.	Re. at 100% Pr.	AP	Pr. at 100% Re.	Re. at 100% Pr.	AP	Pr. at 100% Re.	Re. at 100% Pr.	AP
FastRCNN-AlexNet	94.9%	85.9%	99.2%	84.0%	61.2%	96.8%	69.0%	23.9%	88.4%	67.2%	16.0%	86.2%
	−0.6%	−5.4%	−0.2%	−2.6%	−3.9%	−0.7%	−0.5%	+5.3%	+0.9%	−6.2%	−17.4%	−5.3%
AlexNet	95.5%	91.3%	99.4%	86.6%	65.1%	97.5%	69.5%	18.6%	87.5%	73.4%	33.4%	91.5%
FastRCNN-VGG-M1024	95.5%	84.5%	99.3%	84.0%	48.4%	96.4%	68.8%	35.3%	88.7%	66.2%	14.1%	85.1%
	+1.0%	−0.5%	+0.2%	+2.7%	+2.1%	+0.7%	−1.7%	+10.8%	−0.1%	−8.1%	−13.8%	−6.9%
VGG-M1024	94.5%	85.0%	99.1%	81.3%	46.3%	95.7%	70.5%	24.5%	88.8%	74.3%	27.9%	92.0%