Why not predict the coordinate of the center of bounding box and regress the height and width directly?

[zhangrong1722]

Dear authors,
Thank you for your amazing work firstly. After reading your paper, I found the proposed method uses CNN get the feature map after inputing an image and then predict the top-left and bottom-right keypoint of bounding box, which needs a grouping operation for detected keypoints that may don't work well for crowded people. I wonder whether we can predict the coordinate of the center of bounding box and regress the height and width such that there is no need for following grouping operation.

Thanks.

[ivanpuhachov]

This is what was done in its successor - CenterNet ("Objects as Points")