namepllet / HandOccNet

Offical pytorch implementation of "HandOccNet: Occlusion-Robust 3D Hand Mesh Estimation Network", CVPR 2022.
155 stars 30 forks source link

Fail to test demo on my own images #28

Open oscart01 opened 1 year ago

oscart01 commented 1 year ago

Hello, thanks a lot for your great work. I wanted to test your demo on my own images instead of your original one 'input.png', but found that the snapshot model didn't perform well. Did you select 'input.png' from internet, or did you train the snapshot model on some kind of dataset including this image? And do you know whether it's possible for me to apply your demo to my own data? Thanks!

namepllet commented 10 months ago

I select the 'input.png' from MSCOCO validation set and I trained the model with MSCOCO trainset.

Since I trained the model with limited data, the model may not generalize well.

If you train the model with other kind of datasets such as FreiHAND and YT3D, it will show better results.

yxt7979 commented 10 months ago

I select the 'input.png' from MSCOCO validation set and I trained the model with MSCOCO trainset.

Since I trained the model with limited data, the model may not generalize well.

If you train the model with other kind of datasets such as FreiHAND and YT3D, it will show better results.

@namepllet Hii, and I found that the model retrained on the HO3D dataset works badly on input.png in the folder demo, but the result on HO3D website is normal:

hand_image image

So, actually the online result of HO3D can not really show the ability of the model?

yxt7979 commented 10 months ago

oscartong2001

same question:

截屏2023-11-09 17 14 13