search

mks0601 / I2L-MeshNet_RELEASE

Official PyTorch implementation of "I2L-MeshNet: Image-to-Lixel Prediction Network for Accurate 3D Human Pose and Mesh Estimation from a Single RGB Image", ECCV 2020
MIT License
710 stars 130 forks source link

FreiHAND Training Settings #82

Closed Bozcomlekci closed 2 years ago

Bozcomlekci commented 2 years ago

Greetings,

While training on FreiHAND with train_batch_size=16 and num_gpus=3, I obtained the following results:

PA MPJPE=10.2 mm PA MPVPE=10.3 mm F_aligned@5.0mm=0.555 F_aligned@15.0mm=0.941

The rest of the parameters for training are same with your implementation. I couldn't understand why I obtained the results above instead of the results reported in the paper. Are the default parameter settings implemented applicable for a training on FreiHAND dataset as well? For FreiHAND, what were your training settings such as the values of train_batch_size and num_gpus?

Thanks in advance.

mks0601 commented 2 years ago

Did you use lixel stage? param stage?

Bozcomlekci commented 2 years ago

Did you use lixel stage? param stage?

I used only lixel stage

mks0601 commented 2 years ago

That is weird. Actually, train_batch_size and num_gpus marginally affect the performance. Did you train trainset_3d=['FreiHAND'], trainset_2d=[], testset='FreiHAND' in config.py?

Bozcomlekci commented 2 years ago

That is weird. Actually, train_batch_size and num_gpus marginally affect the performance. Did you train trainset_3d=['FreiHAND'], trainset_2d=[], testset='FreiHAND' in config.py?

Yes, I set them as you describe.

I'll do it all from scratch. If the training settings are same with yours, I guess there is something that I changed and didn't notice.

mks0601 commented 2 years ago

Got it. Let me know if you have some problems.

Bozcomlekci commented 2 years ago

Greetings, It seems that the issue was due to a weight initialization code block that I omit. Doing it from scratch solved the problem. Thanks for the help.