Skip to main content Link Menu Expand (external link) Document Search Copy Copied

Studio + multiple dense

  • When use this: dense cameras, simply calibrate by colmap;
  • Idea: Use texture human body to calibrate the dense rig.
  • Example data: here





Place the data as follows. You can use script to copy the data from raw data:

python3 scripts/preprocess/ <path/to/rawdata> ${root}/ground1f --start 0 --end 1

After this:

├── background1f
│   └── images
├── human1f
│   └── images
└── ground1f
    └── images


Always, we calibrate the camera with human1f which can provide more feature points in the surface. Alternatively, you can use background1f or ground1f.

python3 apps/calibration/ ${root}/human1f ${root}/colmap-human1f --share_camera --colmap ${colmap}

Visualize the results with colmap.

$colmap gui --database_path ${root}/colmap-human1f/database.db --image_path ${root}/colmap-human1f/images --import_path ${root}/colmap-human1f/sparse/0

Detect the chessboard:

python3 apps/calibration/ ${root}/ground1f --out ${root}/ground1f/output --pattern 11,8 --grid 0.06

If the chessboard is so small that can not be detected automaticly. You should manually annotate it.

python3 apps/calibration/ ${root}/ground1f --grid 0.6 0.42 --corner --overwrite
python3 apps/annotation/ ${root}/ground1f --annot chessboard --mode chessboard --pattern 2,2

If you use ChAruco, following this:

python3 apps/calibration/ ${root}/ground1f --mode sparse --show

Align the camera parameters:

python3 apps/calibration/ ${root}/colmap-human1f/sparse/0 ${root}/colmap-align --plane_by_chessboard ${root}/ground1f

Check the camera parameters:

python3 apps/calibration/ ${root}/ground1f --mode cube --out ${root}/colmap-align --show
python3 apps/calibration/ ${root}/ground1f --mode match --out ${root}/colmap-align --show --annot chessboard

Check with Instant-ngp

Currently, ngp only supports cameras with the same camera intrinsics.

# convert the data
python3 apps/calibration/ ${root}/human1f --out ${root}/ngp-human1f --aabb_scale 4 --frame 0
# train and visualize ngp
cd 3rdparty/instant-ngp
# interactively
./build/testbed --scene ${root}/ngp-human1f
# or run in background
python3 scripts/ --scene ${data} --mode nerf --screenshot_transforms ${data}/transforms_novel.json --n_steps 100000 --width 1080 --height 1920 --screenshot_dir ${data}/output --save_snapshot ${data}/ckpt.msgpack