Table 3.

Comparison of fusion timing and methods. In the timing column, det. stands for fusion only before detection (i.e., only the last three auxiliary layers were fused into the backbone); prog. late stands for progressive late fusion, and prog. early is our proposed form of progressive residual fusion. For the fusion methods, concat. stands for fused using concatenation; res. denotes the form of residual learning.

Timing	Method	mAP	mAP (Medium Scale)	mAP (Large Scale)
Det.	concat.	58.51%	9.51%	41.32%
Det.	res.	59.39%	9.56%	42.35%
Prog. late	res.	61.45%	9.99%	45.44%
Prog. early	res.	64.39%	10.93%	46.45%