Skip to content

Latest commit

 

History

History
74 lines (58 loc) · 3.65 KB

README_v3.md

File metadata and controls

74 lines (58 loc) · 3.65 KB
Raw

Yolo_v3_chainer

Yolo v3 implementation by Chainer

各Scale毎に3つのanchorが揃えられている。

YOLO9000: Better, Faster, Stronger link
See the project webpage for more details.

original code: https://github.com/pjreddie/darknet

Download datasets

・COCO
http://cocodataset.org/#download

・VOC
http://host.robots.ox.ac.uk/pascal/VOC/voc2012/#devkit

Converter from darknet to chainer

python to_chainer_converter.py --orig /path/to/original_model --name yolo_v3_chainer

Demo

python demo.py experiments/yolov3_update_608_test.yml --img_path ./data/dog.jpg --thresh 0.20 --nms_thresh 0.3 --save dog
nms: by class
nms_thresh = 0.3
img_thresh = ?

Training

python train.py experiments/yolov3_update_voc_416.yml
nms: by class

Evaluation

python evaluate.py experiments/yolov3_update_416_eval.yml --batchsize 32 --gpu 0
# fps of evaluation includes the time of loading images.
nms: by class
nms_thresh = 0.5
img_thresh = 0.001

Model Comparison from Darknet

Model Dataset darknet map chainer map cfg darknet weight Chainer orig fps chainer fps
YOLOv3 VOC2007+20012 76.8 75.2 link link link **** ****
YOLOv3 544×544 VOC2007+20012 78.6 77.9 link link link
Tiny YOLO VOC2007+20012 link link
YOLOv3 608×608 COCO link link link 58.8 66.6
Tiny YOLO COCO link link
prior version(YOLOv3) COCO link link link

TODO

  • Data loader for imagenet.
  • Training codes for darknet 224×224, 448×448.