Skip to content

Detect macaques with Xilinx KV260! Powered by CenterNet

Notifications You must be signed in to change notification settings

gau-nernst/macaque-detection

Repository files navigation

Macaque detection

Datasets

http://www.pri.kyoto-u.ac.jp/datasets/macaquepose/index.html

Prefix Description Count
PRI Primate Research Institute, Kyoto University (Species: Japanese Macaque) 1641
ZooA Toyama Municipal Family Park Zoo (Species: Japanese Macaque) 3784
ZooB Itozu no Mori Zoological Park (Species: Japanese Macaque) 1312
ZooC Inokashira Park Zoo (Species: Rhesus Macaque) 2747
ZooD Tobu Zoo (Species: Rhesus Macaque) 2461
Other Google Open Images Dataset (Species: Various) 1138
Total Total 13083

To extract bounding boxes from segmentation annotations and export them to COCO format (YOLO is also supported):

python scripts/macaquepose_v1_bboxes.py --data_dir PATH/TO/macaquepose_v1 --output train.json --split train --format coco
python scripts/macaquepose_v1_bboxes.py --data_dir PATH/TO/macaquepose_v1 --output val.json --split val --format coco
python scripts/macaquepose_v1_bboxes.py --data_dir PATH/TO/macaquepose_v1 --output all.json --split all --format coco

Notes:

  • --data_dir should contain annotations.csv file and images folder. pandas and PIL are required.
  • If YOLO format is used, --output should be a subdirectory name.
  • Train split: PRI, ZooA, ZooB, ZooC, ZooD. Val split: Other (Open Images).

Training

Training script is adapted from torchvision. Some default values are changed. We use RetinaNet with ResNet-50 FPN backbone.

python detection/train.py --data-path datasets/macaquepose_v1 -b 16 --epochs 5 --pretrained --amp