search

wenquanlu / HandRefiner

MIT License
755 stars 35 forks source link

MeshGraphormer vs Hamer #35

Open andupotorac opened 5 months ago

andupotorac commented 5 months ago

Hi. I wonder if you guys are familiar with the recently released Hamer, which may be better than MG and how hard it would be to try it out as an alternative? https://geopavlakos.github.io/hamer/

image
wenquanlu commented 4 months ago

Hi, indeed it looks much more robust than MeshGraphormer. Building a preprocessor should be fairly easy if their code can be executed smoothly.

wenquanlu commented 4 months ago

If no one is working on it, I will probably spend some time to push out a new preprocessor based on HaMeR

wenquanlu commented 4 months ago

I have tried the Hamer for handrefiner, it is not suit to estimate malformed hand. Perhaps it's too strong. MeshGraphormer looks more natural, further more Hamer needs much more VRAM. Here is one example: image MeshGraphormer: image Hamer: image

That's very interesting! Thanks for your experimentation, yea perhaps Hamer being so accurate so it even fits the malformed shape, as you said "it is too strong"

andupotorac commented 4 months ago

Thanks guys!

andupotorac commented 4 months ago

I have tried the Hamer for handrefiner, it is not suit to estimate malformed hand. Perhaps it's too strong. MeshGraphormer looks more natural, further more Hamer needs much more VRAM. Here is one example: image MeshGraphormer: image Hamer: image

It would be interesting to see the original image close-ul as well. Did you try with several images of different malformed hands?