kevinlin311tw
commented
5 years ago Thanks for your interest in our repo. I believe there are several misunderstanding.
First of all, people use the COCO2014 Keypoint Dataset instead of the entire COCO2014 dataset for training human pose estimation. Although there are 80K training images in COCO2014 dataset, only some of the images have keypoint annotations.
For COCO2014 Keypoint, there should be only ~45K images that have keypoint annotations, and these annotations are split to train/val sets.
On the other hand, please also note that each image may have more than one person. Given an image that has more than one person, for each person we apply geometric transformation to translate the image making each person located at the center, and thus generate multiple training examples. Thus there is a total of 110K training samples. This is a strategy making the learning process easier. It is common used by many researchers including OpenPose ;)
We also tested the model on COCO Challenging test set, but we think it is difficult to get more insights from the results. This is mainly because a lot of hacking takes place in the challenges. People tend to hack the leader-board using many different approaches such as model ensemble, and the results may not really represent a paper outcome. On the other hand, CMU team uses OpenPose+Person Detection+CPM (a single-person pose estimation for refinement). Though effective, I think reproducing such model combination is already out of the scope of this repo.
H-Liu1997
Hi Kevin, I found in your code the training samples is 110K and val data is about 2K. the training data in offical train2014 is 80K and val2014 is 40k, it seems you using a mix dataset which contain at least 95% of val2014 data. for the accuracy you report, both the mini- val2014 and val 2014 data are at least 95% in your training data, so I don't think the result is so meaningful, have you ever test the result in test-dev2014 or test-dev2017 ? I have read another keras reproduce code in your readme, which you report the model accuracy is 56.7%, they only use the train2014 data. for some other reproduce code, there are also someone using 110K data to train and 2K data for val, Is that because the official openpose using the data set like this? I have read the openpose official code but didn't see that, of course there maybe some misunderstanding of some part. but totally, I still confuse that there didn't have some overfitting about this result? sorry to bother you and looking forward to your reply