JGRCAN: A Visual Question Answering Co-Attention Network via Joint Grid-Region Features

Table 1

Convolution dimensional transformation process.

	Input	Conv + MaxPool	RB^1 × 3^2	RB × 4	RB × 3	RB × 23	AvgPool
Layer structure	—	3 × 3, Stride: 2					2 × 2, Stride: 2
Shape	3 × 448 × 448	64 × 112 × 112	256 × 112 × 112	512 × 56 × 56	1024 × 28 × 28	2048 × 14 × 14	2048 × 7 × 7

RB is the abbreviation of residual block; RB × 3 means that three residual blocks are used.