Does misalignment of heatmap prediction harm?

[lck1201]

Hi, In coco dataloader of training stage, you generate heatmap by sigma=7.0 regardless of its scale. And in evaluation, you use multi-scale images as input. To my understand, heatmap is trained in sigma=7, so images of all scales tend to predict heatmap of sigma=7. Then you just rescale and average all the heatmaps, which means gaussian shape of small heatmap will be enlarged, and the one of big heatmap will shrink.

Does this multi-scale heatmap misalignment harm the prediction？Have you ever consider taking scale in to account, and generate heatmap based on image scale?