Jia Li
@dong-x16 Could you please share your results on mscoco 2017 test-dev subset without refinement? I need to compare the difference. Thank you!
Another issue: Is the joint xyz coordination in get h5dy data code also normalized to [-1,1] in your code just as the depth do?
PS. Both of you aim to estimate the distance in meters.
Same problem. At first, I can use apex package well on level O1, but I add Stochastic weight averaging, this kind of problem appears.
When I try to train the network with frozen BN layers. I can continue to reload and train the net if I convert the BN layers to half manually.
In my case, I exactly switch from train() in the training loop and free the BN just after the train() in the training loop. I'll share my post a few...
Hi, do you use batch testing (let's say batch_size=32) during evaluating speed?
Then, what is the difference whether we use bias or not?
Hi everyone. How to handle the joints of instance which is crowded in annotation？ It seems that the keypoints of these areas are not annotated.
OK， got it. We will ignore an area if there is at least one instance is not annotated.