Step by step 3d multianimal visualisation with DeepLabCut

This repo is a shorter documentation of 3D DeepLabCut

Pull a docker image of DLC:

docker pull deeplabcut/deeplabcut:2.3.5-base-cuda11.7.1-cudnn8-runtime-ubuntu20.04-latest

Run the dlc container & run python inside the container:

Enter the container:

./run_dlc.sh

Correct the version of numpy, otherwise it messes up with a library at triangulate

pip install numpy=1.22

Enter the python environment in the container:

python3

Import the DLC module in python:

import deeplabcut

Create the calibration images:

deeplabcut.calibrate_cameras("/host/analysis_cam2_cam3/config.yaml", cbrow=3, cbcol=6, calibrate=False, alpha=0.5)

Then rerun with

deeplabcut.calibrate_cameras("/host/analysis_cam2_cam3/config.yaml", cbrow=3, cbcol=6, calibrate=True, alpha=0.5)

Check for undistortion

deeplabcut.check_undistortion("/host/analysis_cam2_cam3/config.yaml", cbrow=3, cbcol=6)

Triangulate:

⚠️ You will get the error UnboundLocalError: local variable 'putativecam2name' referenced before assignment IF your videos do not have a suffix or prefix. So far we have been working with the names camera-01 and camera-04 because we calibrate the CAMERAS and not the VIDEOS.

deeplabcut.triangulate first predicts the pose in the videos (hence why we need a suffix or prefix). And then create triangulation files for each pair of videos.

deeplabcut.triangulate("/host/analysis_cam2_cam3/config.yaml", "/host/analysis_cam1_cam4/videos/", filterpredictions=True, gputouse=1,save_as_csv=True)

It's also possible to feed in specific videos we want to analyze instead of the single path.

Create the 3d video:

deeplabcut.create_labeled_video_3d("/host/analysis_cam1_cam4/config.yaml", ["/host/analysis_cam1_cam4/videos"], start=4800, end=5300, videotype=".avi", xlim=(0,10), ylim=(-30,-10), zlim=(10,30))

NINAnor/3D_deeplabcut