. Author manuscript; available in PMC: 2019 Sep 1.

Published in final edited form as: Med Image Comput Comput Assist Interv. 2018 Sep 26;11071:921–929. doi: 10.1007/978-3-030-00934-2_102

Table 1.

The performance of different PL localizers on the test set via 10-fold cross validation. All results (%) are in the format of median [first, third quartile]. APx at three SOKS thresholds (x ∈ {0.3, 0.5, 0.75}) are reported. mAP is the primary metrics. Models are named in the format of A-B, where A is its generic architecture (DN or UN) and B its layer aggregation structure (DS or DS-FCN or DS-IDA)

Model	Backbone	mAP	AP_0.3	AP_0.5	AP_0.75
DN-DS	ResNet18	22.8 [20.2, 26.5]	33.0 [30.3, 36.3]	29.6 [27.1, 33.8]	19.7 [16.9, 23.9]
DN-DS-FCN		28.7 [22.0, 30.0]	42.7 [34.5, 43.6]	37.5 [30.6, 39.3]	24.0 [16.9, 25.7]
DN-DS-IDA		29.7 [25.3, 34.8]	38.6 [33.8, 46.0]	36.3 [31.0, 43.9]	28.5 [24.1, 32.5]
UN-DS	VGG16	32.6 [24.1, 37.5]	41.4 [35.2, 47.2]	39.6 [31.4, 44.1]	31.3 [22.0, 36.5]
UN-DS-FCN		32.2 [28.4, 37.4]	42.3 [39.7, 46.0]	40.2 [35.7, 44.1]	31.0 [24.8, 36.8]
UN-DS-IDA		35.7 [28.4, 40.7]	44.7 [40.9, 50.1]	42.3 [36.5, 48.5]	35.3 [26.4, 37.8]