Scale Adaptive Feature Pyramid Networks for 2D Object Detection

Table 3

Accuracies in mAP [%] of the faster R-CNN [5] without FPN, with (original) FPN, and proposed SAFPN.

	W/O FPN	FPN	SAFPN
			L0.5	L1	L2	(L2)^2	L∞
Faster R-CNN	76.4	78.4	78.9	79.5	79.5	79.9	76.5
RetinaNet		77.6	76.1	77.2	78.3	78.5	76.4

Accuracies for RetinaNet [10] are also shown with the FPN and SAFPN. Training is done using the Pascal VOC 2007 + 2012 joint trainval dataset.