question about Figure 3 in the paper

[yijiudd]

Dear author, I have a question about this figure, when you do the experiments on gradual shrinking, the improved AP when shrinking the depth means on the low-computational region, or all regions on an image compared to the HigherHRnet.I am curious why decreasing the branch will have better accuracy. And what does the low-computational region exactly mean? Thanks for your time.