search

Juyong / Fast_RNRR

Source code for the paper "Quasi-Newton Solver for Robust Non-Rigid Registration" (CVPR2020 Oral).
264 stars 51 forks source link

face registration result is very bad #3

Closed ZHANG-SHI-CHANG closed 4 years ago

ZHANG-SHI-CHANG commented 4 years ago

the result in face registration is very bad, like follow, 134 landmarks is used: source: image target: image nricp result: image fast rnrr result: image

Juyong commented 4 years ago

Hi, how do you construct the correspondence between source and target models? If I understand correct, they have the same topologies. Besides, how do you set the weights?

ZHANG-SHI-CHANG commented 4 years ago

Hi, how do you construct the correspondence between source and target models? If I understand correct, they have the same topologies. Besides, how do you set the weights?

yes, they have the same topologies, but i just use 134 landmarks, because i want to use it between different topologies. And i didn't modify the code about weights

Juyong commented 4 years ago

Can you send the test models and the labeled landmarks to us via juyong@ustc.edu.cn?

ZHANG-SHI-CHANG commented 4 years ago

Can you send the test models and the labeled landmarks to us via juyong@ustc.edu.cn?

👌，thx~

yaoyx689 commented 4 years ago

Hi, firstly, our method depends on geodesic calculation when sampling nodes, the geodesic code we used currently only supports closed mesh calculations, so maybe you need to fill holes first and then run it or change to a more suitable geodesic calculation method(in nodeSampler.cpp and geodesic.cpp). Secondly, the weights of regularization term and rotation term will greatly affect the results, you can adjust(decrease) them (paras.alpha and paras.beta in main.cpp) to make it more suitable for your models.

ZHANG-SHI-CHANG commented 4 years ago

Hi, firstly, our method depends on geodesic calculation when sampling nodes, the geodesic code we used currently only supports closed mesh calculations, so maybe you need to fill holes first and then run it or change to a more suitable geodesic calculation method(in nodeSampler.cpp and geodesic.cpp). Secondly, the weights of regularization term and rotation term will greatly affect the results, you can adjust(decrease) them (paras.alpha and paras.beta in main.cpp) to make it more suitable for your models.

ok, i get it

X-niper commented 1 year ago

Hi, firstly, our method depends on geodesic calculation when sampling nodes, the geodesic code we used currently only supports closed mesh calculations, so maybe you need to fill holes first and then run it or change to a more suitable geodesic calculation method(in nodeSampler.cpp and geodesic.cpp). Secondly, the weights of regularization term and rotation term will greatly affect the results, you can adjust(decrease) them (paras.alpha and paras.beta in main.cpp) to make it more suitable for your models.

ok, i get it

Hi， Mr Zhang，

Have you finally gotten good results on face registration?