Question about the different image size in training and testing

hackiey / keypoints

A pytorch implementation of Google's paper <Towards Accurate Multi-person Pose Estimation in the Wild>

MIT License

102 stars 25 forks source link

Question about the different image size in training and testing #7

Open HowieMa opened 4 years ago

HowieMa commented 4 years ago

Hi, thank you for providing this excellent code. I have some questions about the different image size at training and testing. In training, you calculate loss based on the upsampled output, which has the same image size as the input image. However, in testing(prediction), you calculate the keypoints based on the results before Upsampling. I was wondering why you do this in testing time? For efficiency or something else? Look forward to your reply, thank you!

hackiey commented 4 years ago

This repo has been for a long time, and I have a bit of trouble remembering these meanings.

Actually I think the results before upsampling is just for analysis, and the real results is after upsampling, and upsampling has no parameters, so the results are consistent in a way.