search

justanhduc / graphx-conv

Official implementation of GraphX-Convolution
https://justanhduc.github.io/2019/09/29/GraphX-Convolution.html
MIT License
62 stars 17 forks source link

Trained on my own data #10

Closed brian220 closed 3 years ago

brian220 commented 3 years ago

Hi, thanks for the nice implementation.

I have trained your model on my own images and point clouds, And I found that the reconstruction results is different when I use different point clouds dataset. For example, when I use the dataset which the chair point clouds stand on the x-y plane will bring the better results (more details) then which stand on y-z plane. It seems that the orientation of ground truth point clouds may influenced the reconstruction results. Should I adjust the code according to different orientation of ground truth data? It would be very helpful if you could give me some suggestions.

Thank you very much!

justanhduc commented 3 years ago

Hi @brian220. Nice observation. I can just have an educated guess for this. Probably the deficiency comes from this, which is a heuristic initiation. This initial point cloud is to ensure that when collapse z, it will cover the whole image plane, which is x-y. Then, theoretically, you should obtain the same result if initializing from isotropic shapes like sphere. Actually I tried it before but it gave poorer results, but if you need something robust to orientation, it should be the one to go to.

Code bonus:

def sample_spherical(n_points):
    vec = np.random.rand(n_points, 3) * 2. - 1.
    vec /= np.linalg.norm(vec, axis=1, keepdims=True)
    pc = vec * .3 + np.array([[6.462339e-04,  9.615256e-04, -7.909229e-01]])
    return pc.astype('float32')
brian220 commented 3 years ago

Hi, @justanhduc
Thanks for your reply, which let me have a more profound understanding of the network. I will try with different initial to check the difference.

Thank you very much!