Do you test your model on the images in the wild？ about handmesh HOT 13 CLOSED

Yes. Our model has a nice generalization. We have used it for online applications. Note that a small model size leads to better generalization.

from handmesh.

Thank you, that's great! And I also puzzle about whether your 3d mesh is the same to the MANO or not. Hope for your reply.

from handmesh.

Yes, the topology is the same as MANO.

from handmesh.

好的，谢谢大佬。nice work~膜拜一下

from handmesh.

Hi, xy:
I try your demo on my hands and get the following results, it's there any thing wrong? I use the camera_param of 20_img as my camera_param.

from handmesh.

Hi, camera_param is used to compute camera-space information, and the hand pose is not related to K. I have tried the first sample. CMR-PG has the below results. Because of the lighting change, it brings errors.

So, I have released CMR-G. It has smaller model size and better generalization, and we train it with FreiHAND + other data. The model has the below results. However, the silhouette predicted by this model contains the wrist, so you would want to use the 2D registration only for mesh-image alignment, i.e., set poly=None in registration.

Announce in advance: we will release a new version of this method working on mobile devices in 3 months.

from handmesh.

Thank you for your reply. So is there any differences between training CMR-G with others model?

from handmesh.

All models have the same training setting.

from handmesh.

ok, I'll try to train a model by myself. THx for your help.

from handmesh.

Hi, I also have the same problem when I run demo on my own images, Do you find how to change the settings or something to get the good results?

from handmesh.

Hi, I also have the same problem when I run demo on my own images, Do you find how to change the settings or something to get the good results?

Hi, thanks for your question. Please wait for days. I will update this repo and release the model for real-world test.

from handmesh.

Hi, I also have the same problem when I run demo on my own images, Do you find how to change the settings or something to get the good results?

Hi, thanks for your question. Please wait for days. I will update this repo and release the model for real-world test.

Thanks for your reply, so the results for real-world example above is generated by the other trained-model instead of model which was released?

from handmesh.

Yes. Our model has a nice generalization. We have used it for online applications. Note that a small model size leads to better generalization.

@SeanChenxy great work! Could you please tell me what inference framework for your online application ? Thanks !

from handmesh.