AlphaPose
AlphaPose copied to clipboard
MVIG-SJTU
FastSEResNet and DUC layer
I found AlphaPose is totally different from RMPE, it is faster and more accurate but without the key features of SSTN, Parallel SPPE and etc. AlphaPose is simpler and the SPPE is use FastSEResNet and DUC layer, could you give the papers where these layers come from? and why they work well?
Hi, thanks for your interest. In fact we are about to write the paper, but not yet 😂
@Fang-Haoshu I try to train the model using gluoncv code with 1 gpu, but the network didn't converge and val accuracy is quite low, Is the gluoncv version same with this mxnet branch? I found the training data was a little different from gluoncv's,
- trainval2017 is train2017 + val2017?
- annot_coco.h5 is the same with annot.h5 in the master branch?
- why use person_keypoints_val2014.json not 2017?
The gluoncv code should work properly. We have worked for a long time on that PR. It's a bit different.
Hi, thanks for your interest. In fact we are about to write the paper, but not yet 😂
Just wanted to know whether the paper had been published yet.
Hello @7oud, did you said that AlphaPose doesn't use SSTN, Parallel SPPE, etc. Do you know the full pipeline of this AlphaPose? What technique does it use? Maybe you can help me? Sorry if bothering you all @Jeff-sjtu @Fang-Haoshu
Hi, the paper is public at http://arxiv.org/abs/2211.03375 Thanks for your interest!