Scenes

Each scene contains 2 folders:

1. data
2. camera_poses

data is the true output for our dataset in terms of per-frame data. The other stuff is what we use to generate data.

data contains frames of format 00001_color.jpg, 00001_depth.png, 00001_label.png, 00001_meta.txt.

• 00001_color.jpg: color image (3 channel 8 bit) (comes from camera)
• 00001_depth.png: depth image (16 bit 1 channel) (comes from camera, aligned to color)
• 00001_label.png: ground truth semantic segmentation (1 channel 8 bit). We generate this using a virtual camera.
• 00001_meta.json: ground truth poses of objects in the scene. We calculate this using OptiTrack.

camera_poses contains data pertaining to the camera poses:

1. camera_poses.csv is raw optitrack pose data
2. camera_poses_cleaned.csv is cleaned optitrack pose data
3. camera_poses_synchronized.csv:
   • this contains the virtual camera pose in optitrack coordinates for each frame
   • generated by CameraPoseSynchronizer
   • used by other parts of the codebase
   • not shipped as part of the dataset

Each scene also contains a file: scene_meta.yaml:

• this contains scene level metadata
  • camera intrinsics
  • camera depth scale
  • which objects are in this scene
  • random stuff like time of day?