YOLOv3

minimal implementation of YOLOv3 in pytorch.

run the following command to start training on the default setting.

python3 yolov3/train.py

example result (y-axis is log scaled)

mAP: 0.5328507833182812
recall: 0.5872999429702759
precision: 0.4169691205024719

dataset (PASCAL VOC in YOLO format [cx, cy, w, h])

download HERE

or

run download_pascal_voc.py

reference

original code from https://github.com/aladdinpersson/Machine-Learning-Collection/tree/master/ML/Pytorch/object_detection/YOLOv3

TODO

• [ ] fix mAP error (current implementation introduce minor numerical error compared to torchmetric)
• [ ] add coco.py under yolov3/datasets
• [ ] more augmentation methods (e.g. pastein, mosaic, mixup, ...)
• [ ] experiment simOTA