README.md

October 25, 2022 ยท View on GitHub

Preparing depth, rgb and mask data

  • Follow the data format as our prepared example
  • All images must be in 'jpg' or 'png' format
  • Count from zero, e.g. the first frame is '00000.png'
  • The shape of depth is (H, W), and the shape of rgb is (H, W, 3)
  • The shape of mask is (H, W, 3). Pixel in the object is [255, 255, 255], out of object is [0, 0, 0]
  • Important: Include a file for intrinsic matrix as 'intrinsic.txt' (containing focal length and center at least) in [data_path_]

Compiling Robust-ICP

  • Clone Robust-ICP in current dir
  • Compile it and make sure FRICP in path ./Fast-Robust-ICP/build/FRICP, named as [FRICP_path_]

Step 1: Registrating depth maps with Robust-ICP

  • Run
python step1.py --depth_scale [depth_scale_] --data_path [data_path_] --registration_alg_path [FRICP_path_]

e.g.

python step1.py --depth_scale 1000. --data_path '../datasets/kfusion_frog/' --registration_alg_path './Fast-Robust-ICP/build/FRICP'

Step 2: Checking registrated depth maps scaled into unit ball

  • Put unit ball and all depth maps from [data_path_]/intermediate/colored_pointclouds/ into MeshLab
  • Delete invalid points which are not on the object surface
  • Save the result as [data_path_]/intermediate/ref.xyz

Logo

Step 3: Obtaining camera parameters

  • Run
python step3.py --depth_scale [depth_scale_] --data_path [data_path_] --object_scale [object_scale_]

e.g.

python step3.py --depth_scale 1000. --data_path '../datasets/kfusion_frog/' --object_scale 1.03
  • Put unit ball and all depth maps from [data_path_]/intermediate/colored_pointclouds/ into MeshLab again
  • Make sure invalid points out of ball
  • Make sure valid points on the object surface fill in the ball as much as possible

Tips: Set [object_scale_] to make sure object sequence scaled into the unit ball. Larger it is, smaller scaled object will be.