Is it possible to train a high-quality model that predicts the 3D coordinates of the animal's body surface from a photograph, without the corresponding DensePose markup?
Researchers from Facebook Artificial Intelligence Research raised this issue at the CVPR 2020 machine vision conference.
About the DensePose task
DensePose-COCO , (). COCO 2014.
:
bounding boxes ,
pixel-perfect foreground-background ,
32 , ,
(c, u, v) , c — , u , v — .
:
( ) 3D , .. 2D SMPL . (c, u, v) .
5 50 COCO 2014.
Mask-RCNN 3D .
- . , . :
DensePose-COCO COCO Dataset , 3D . . , , , .
DensePose Average Precision = 34.9. , = 46.8. , 0 100. knowledge transferring?
, ( , ). SMPL 3D .
, , .
Show me the code!
As is usually the case with articles published by eminent scientific groups like FAIR - they are accompanied by a code. For both articles, it is available inside the official detectron2 repository on GitHub .
Early code from the first article, written using the first version of detectron, which is based on Caffe2, can also be found on GitHub.