training_data.md

Training Data

We provide a few suggestions on how to setup training data for GRIT. We focus here on how to get data and annotations from datasets used in GRIT that you are allowed to train on.

Surface Normals

Many of the surface normals in GRIT have been created from some other type of 3D data (e.g. depth, point clouds). Therefore, to get surface normals for the images in the GRIT dataset, you either need to generate them yourself or download the normals from someone else who has already done this.

ScanNet

For ScanNet images, we use surface normals created by the FrameNet folks. Use the provided download script:

• Downloads a bunch of zips from FrameNet. These are listed in scannet-frame-links.txt
  • http://download.cs.stanford.edu/orion/framenet/scannet-frame/scene0016_02.zip
  • http://download.cs.stanford.edu/orion/framenet/scannet-frame/scene0631_02.zip
  • http://download.cs.stanford.edu/orion/framenet/scannet-frame/scene0748_00.zip
  • … and so on
• The zips contains frames from the scene (at intervals of 10)
  • scene0016_02 has frames range(0,930,10)
  • scene0575_00 has frames range(0,2930,10)
• Each frame has 6 variants saved as .png images
  • scannet-frames/scene0016_02/frame-000000-color.png
  • scannet-frames/scene0016_02/frame-000000-normal.png
  • scannet-frames/scene0016_02/frame-000000-orient-X.png
  • scannet-frames/scene0016_02/frame-000000-orient-Y.png
  • scannet-frames/scene0016_02/frame-000000-orient-mask.png
  • scannet-frames/scene0016_02/frame-000000-orient.png
• You only need to keep the color.png and normal.png for this task
• Remember not to train on images in the GRIT ablation or test