search

DLR-RM / AugmentedAutoencoder

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

Questions about generating datasets #103

Closed JZhao12 closed 2 years ago

JZhao12 commented 2 years ago

Hi, Martin

sorry to disturb you, may I ask what kinds of rendering software you used to generate your database from different views in your method? I would like to generate some LineMod images like this:

001276

Specifically, I would like to first randomly create multiple rotation matrices, then apply these rotation matrices to objects' 3D models, and then get the rotated images. I noticed that your department has a rendering software called BlenderPro, is this convenient to generate images as I described above?

Thank you very much! Jianyu

MartinSmeyer commented 2 years ago

Hi,

I created my own renderer in OpenGL, it is included under auto_pose/meshrenderer. However these days I would rely on pyrender for such simple rasterized images.

You can also use our software BlenderProc to do that if you want more photorealistic renderings using ray tracing. https://github.com/DLR-RM/BlenderProc/tree/main/examples/basics/camera_object_pose

To obtain strong generalization performance you can also train on the BlenderProc scenes that we created for the BOP challenge. https://github.com/DLR-RM/BlenderProc/blob/main/README_BlenderProc4BOP.md