Figure 3.

The D-WASP disentangled waterfall module. The feature sizes are denoted by the two spatial dimensions, followed by the channel dimension. The inputs are 32, 64, 128, and 256 feature maps from all four levels of the HRNet backbone, as illustrated in Figure 2, and low-level features from the initial layers of the framework. The module processes the backbone features at different rates of dilation in a waterfall fashion and outputs both the keypoints and offset heatmaps for each person instance.