search

DLR-RM / AugmentedAutoencoder

Official Code: Implicit 3D Orientation Learning for 6D Object Detection from RGB Images
MIT License
338 stars 97 forks source link

'not enough visible points' when testing with TejaniDB milk object #39

Closed HsiaoTsan closed 4 years ago

HsiaoTsan commented 5 years ago

Thank you for your brilliant masterpiece and making it open-sourced. I have successfully trained AAE on TejaniDB object "milk", when I run ae_eval, the code warns me not enough visible points, then the image shown does not render the estimation object. I checked the code it is in auto_pose/eval/icp_utils.py, line 265.

    if len(real_depth_pts) < len(synthetic_pts)/8.:
        print 'not enough visible points'

It seems that the number of real depth points is not enough. I have changed the if statement to len(real_depth_pts) < len(synthetic_pts)/800., but the warning still exists. What should I do? Thank you in advance!

MartinSmeyer commented 5 years ago

Hi HsiaoTsan, thank you and sorry for my late reply. This is within the iterative depth refinement (ICP) and happens when they are not enough matches between the rendered model and the scene. Does it occur every time? How is the output if you run your code without the ICP, i.e. setting icp = False in you eval config? What is the value in your training cfg for vertex_scale?

HsiaoTsan commented 5 years ago

Hi Martin, thank you for your reply and sorry I'm late. It happens to all TejaniDB object 'milk', no estimated pose will be rendered into the scene image. It happens every time. If I set the icp = False the output is as normal as TLESS datasets. the vertex_scale is 1 in the training cfg, as the point cloud .ply file is already in millimeter.

I have not worked in TejaniDB dataset for a while, you may close this issue if you want to.

Thank you.