Official Implementation of Tracking Grow-Finish Pigs Across Large Pens Using Multiple Cameras
CVPR2021 CV4Animals Workshop Poster
The dataset can be found at this link.
The videos were acquired at the Imported Swine Research Lab (ISRL) at UIUC. The deployment video can be found here. It was annotated for grountruth global identities and bounding boxes using this MATLAB Tool.
deepsort-tracking: Contains code for detecting and tracking pigs using YOLOv4 and DeepSORT
data/homography: Contains pickled homography matrices for both the pens
camera.py: Detects and tracks pigs using the trained model in DeepSORT and YOLOv4
manager.py: Main file which uses the homography matrices to assign global identities
Download the dataset from this link and place multicam-dataset
folder in data/
. Note that we have already trained the model and extracted the output of DeepSORT into JSON files. You can find the pretrained checkpoint here
- Run
export DARKNET_PATH=./deepsort-tracking/yolov4/
in terminal. - Run any one of the commands from commands.txt, for instance: python3 manager.py --av data/multicam-dataset/0/0-Pen_B.mp4 --cv data/multicam-dataset/0/0-Ceiling_Cam.mp4 --cj data/multicam-dataset/0/0-Ceiling_Cam.json --aj data/multicam-dataset/0/0-Pen_B.json --cl 457
We are currently working on building action recognition models for pig behavior using ethograms. We can currently estimate the time spent by pigs near drinkers and feeders based on their proximity.