What parameters were used on MPII dataset?

[yw155]

Hi @ZheC, I use you provided 'MPI.json' file to extract 343 images as the validation set to test the accuracy of you provided MPII-iter146000 model. The network input size is set as 368. The scale search uses 3 scales, which are 0.7, 1 and 1.3. The obtained accuracy is only 50.4 and much lower than the accuracy mentioned in the paper (79.1). I am not sure if it is the reason of parameter setting. Could you tell me what parameters were used on MPII val set? Thanks.

[ZheC]

The parameters I used is in:
https://github.com/ZheC/Realtime_Multi-Person_Pose_Estimation/blob/master/testing/config.m#L66-L85

50.4 seems to be a quite low number, probably you should plot the prediction and GT pose to check if anything is obviously wrong.

[yw155]

Hi @ZheC, I use all the default parameters and model to generate the predictions on the val set of 343 images. Below are the detected pose results. I checked all the results and did not see the result is obviously inaccurate. So according to your experience, what is the possible reason of such a low accuracy?

[ZheC]

I am closing this issue because you are posting the issues in a new thread (#172)