PersonLab

ECCV 2018

box-free bottom-up approach, pose estimation, instance segmentation

individual keypoints and relative displacements and part-induced geometric embedding descriptor

to associate semantic person pixels, person instance

achives coco tes-dev keypoint average precision of 0.665 using single-scale inference

0.687 using muti-scale inference, coco instance segmentation task average precision of 0.417.

 

17 keypoints face and body parts in the coco dataset.

instance-agnostic fashion, all visible keypoints belonging to any person in the image.

a disk of radius R centered around y. let yj,k be the 2-D position  of the k-th keypoint of the j-th person instance.

deliberately opted for a disk radius

heat map loss as the average logistic loss 

person crowd areas and small scale person segments

short-range offset vectors. back-propagateing the errors only at the positions in the keypoint disks.

we divide the errors in the short-range offsets by the radius R pixels in orderto normalize them and make their 

dynamic range commensurate with the heatmap classification loss.