Designing Compact Convolutional Filters for Lightweight Human Pose Estimation

Figure 3

The comparison of two different upsampling methods. (a) The traditional transposed convolution, which has a large computational overhead. (b) The proposed lightweight upsampling block, which includes depthwise transposed convolution operation, pointwise convolution operation, and attention . In (a), the features are amplified directly by transposed convolution. In (b), we first use point convolution to expand the number of channels of the feature so that the number of channels goes from to and then use the depthwise transposed convolution to generate high-resolution feature maps. Finally, we use a point convolution to change the number of channels to and the attention mechanism to make feature map stronger.

(a)

(b)