search

mks0601 / Integral-Human-Pose-Regression-for-3D-Human-Pose-Estimation

PyTorch implementation of "Integral Human Pose Regression", ECCV 2018
194 stars 25 forks source link

Large value of MPJPE using the trained model. #11

Closed xuyanyu-shh closed 5 years ago

xuyanyu-shh commented 5 years ago

Thanks for your work. Very thanks.

I have generated the H36M data and downloaded your provided trained model. However, the value of MPJPE is very large, using your trained model.

integral_loss

As you said, 'set testing set in config.py'. Is there something modified? I just use your provided config without modification.

Thanks very much! ")

xuyanyu-shh commented 5 years ago

In addition, the following is my current configuration: Python 3.6.7 Pytorch 0.4.1 CUDA 9.0

I will update my pytorch and test the model again.

xuyanyu-shh commented 5 years ago

Unfortunately, I tested the provided model under pytorch 1.0 version. The value of MPJPE does not change.

mks0601 commented 5 years ago

I tested using code and model of this repo, but there is no problem. Below is the result of mine.

Evaluation start...

Protocol #1 error (PA MPJPE) >> 41.46
Protocol #2 error (MPJPE) >> 53.86

Protocol #1 error (PA MPJPE) for each action: 
Directions: 38.99
Discussion: 38.64
Eating: 44.08
Greeting: 42.53
Phoning: 40.63
Posing: 35.26
Purchases: 38.16
Sitting: 49.94
SittingDown: 59.35
Smoking: 41.00
Photo: 46.11
Waiting: 37.63
Walking: 30.30
WalkDog: 40.81
WalkTogether: 35.51

Protocol #2 error (MPJPE) for each action: 
Directions: 50.82
Discussion: 52.29
Eating: 54.79
Greeting: 57.91
Phoning: 52.76
Posing: 47.03
Purchases: 52.10
Sitting: 62.04
SittingDown: 73.65
Smoking: 52.63
Photo: 58.28
Waiting: 50.44
Walking: 40.94
WalkDog: 54.11
WalkTogether: 45.08
mks0601 commented 5 years ago

Close the issue. Ask me you have any questions :)