Include PyTorch and Darknet implementation for VisDrone dataset.
git clone https://github.com/zhaobaiyu/visdrone.git <path-to-repository>
yolo.md 包含yolo的一些资料
dataset.md 关于pascal VOC、MS COCO以及VisDrone数据集的结构和格式
git clone https://github.com/pjreddie/darknet.git <path-to-darknet>
cd <path-to-darknet>
wget https://pjreddie.com/media/files/yolov3.weights
wget https://pjreddie.com/media/files/yolov3-tiny.weights
# cuda or opencv option needs to modify Makefile, see the website above
make
visdrone_vid_transform.py 对于VisDrone中的视频目标检测问题，该脚本将数据转换为Yolov3 Darknet实现所需格式，目录结构和Pascal VOC相似
visdrone_det_transform.py 对于VisDrone中的图片目标检测问题，该脚本将数据转换为Yolov3 Darknet实现所需格式
run the command below twice, with respect to VisDrone2018-VID-train and VisDrone2018-VID-val
cd <path-to-dataset>
cp <path-to-repository>/darknet/scripts/visdrone_vid_transform.py .
python visdrone_vid_transform.py
cd <path-to-darknet>
cp <path-to-repository>/darknet/cfg/visdrone.data cfg/
# visdrone_det.data for detection problem
cp <patn-to-repository>/darknet/cfg/yolov3-visdrone.cfg cfg/
cp <path-to-repository>/darknet/data/visdrone.names
modify<path-to-darknet>/cfg/visdrone.data：
classes= 10
train = <path-to-visdrone-train-dataset>/images.txt
valid = <path-to-visdrone-val-dataset>/images.txt
names = data/voc.names
backup = backup
modify<path-to-darknet>/cfg/yolov3-visdrone.cfg
[net]
# Testing
# batch=1
# subdivisions=1
# Training
batch=64
subdivisions=8
....
cd <path-to-darknet>
./darknet detector train cfg/visdrone.data cfg/yolov3-visdrone.cfg darknet53.conv.74
PyTorch implementation from scratch
代码暂未写完。