The effect of object reconstruction

[yys-abu]

Hello,

Thank you for the great work(s)!

I set up an environment on WSL, but perhaps due to a flaw in WSL, I couldn't use EGL to render. So, I use OSMESA to render and comment the command:

if 'PYOPENGL-PLATFORM' not in os. environment:
os.environ['PYOPENGL_PLATFORM'] = 'egl'
Revise to:
os.environ['PYOPENGL_PLATFORM'] = 'osmesa'

The demo can run smoothly, but an error occurred:
cam_view = renderer.render_rgba(verts, pointclouds_up + palm + objtrans, cam_t=cam_t, render_res=(224, 224), is_right=True, **misc_args)
The size of the cam-view obtained in this step is (224，224，3) instead of (224，224，4)

so I manually added an alpha_channel=np.ones((cam_view.shape[0], cam_view.shape[1], 1))
cam_view = np.concatenate([cam_view, cam_view], axis=-1)
The final result of visualizing the test2.json file is:

The reconstructed object point cloud is not as good as described in the paper. Is it due to issues with cam-view and not using EGl rendering, or is the result just like this.

[zerchen]

Hi @yys-abu,

Thank you for your interests and kind words. Sorry, I currently do not have a computer with GPUs and cannot test the omesa on my side. Therefore, I suggest to compare the quality of point clouds with our hugging face demo. By comparison, you could find out whether the issue comes from the rendering part in WSL.

Best,
Zerui

[yys-abu]

Hi @yys-abu,

Thank you for your interests and kind words. Sorry, I currently do not have a computer with GPUs and cannot test the omesa on my side. Therefore, I suggest to compare the quality of point clouds with our hugging face demo. By comparison, you could find out whether the issue comes from the rendering part in WSL.

Best, Zerui

Thank you for your reply. Is it true that hort doesn't perform well in reconstructing larger square boxes.