search

MVIG-SJTU / AlphaPose

Real-Time and Accurate Full-Body Multi-Person Pose Estimation&Tracking System
http://mvig.org/research/alphapose.html
Other
7.94k stars 1.97k forks source link

Training SPPE #425

Closed DonaldKam closed 4 years ago

DonaldKam commented 4 years ago

Glad to find the amazing Alphapose and i try the codes in “train_sppe” and succeed in running the codes. But i am confused about the displayed information "acc". What is the "acc" used to evaluate? Keypoint detection? If so, where is the human detector in “train_sppe” as i think codes in “train_sppe” are only for SPPE, without human detector? If “train_sppe” does not contain human detector，how can it performe keypoint detection? Another Q，as is described in Readme， annot_coco.h5 and person_keypoints_val2014.json are needed during training SPPE，what are they used for, respectively？ If they are labels for training and validation，why not both in h5 or json format？ Looking forward to the reply.

thancaocuong commented 4 years ago

They generate sample box from annotation to simulate result from detector. So they dont need detector backbone when train sppe.

thancaocuong commented 4 years ago

Btw, I used coco api to load data from annotation, dont need h5 file 😂😂

Fang-Haoshu commented 4 years ago

Hi, we update the training code for coco. It's more user-friendly now. You can check out the master branch for more details.