1、修改makefile：GPU=1，CUDNN=1,OPENCV=1,LIBSO=1;
2、make 或make -j16;
3、数据集准备：VOCdevkit——>VOC2008;
4、运行xml2txt.py,接着运行voc_label.py;
5、修改yolov4-tiny-custom_2008.cfg:
max_batches = 500200 # 训练达到max_batches停止，具体数值为classes2000，但是最小数值不能小于自己数据集图片的数量以及最小值不能小于6000
steps=400000,450000 # 步长 数值为max_batches的80%和90%
找yolo,2个，修改classes=类别数，修改yolo前面的filters数目，filters=3(classes+5)；
6、修改voc.names
7、修改voc.data
8、训练：./darknet detector train data/voc.data yolov4-tiny-custom_2008.cfg yolov4-tiny.conv.29 -map
9、暂停后继续训练：./darknet detector train data/voc.data cfg/yolov4-tiny.cfg backup/yolov4-tiny-custom_last.weights -map -gpus 0,1
10、图片测试：./darknet detector test data/voc.data yolov4-tiny-custom_2008.cfg backup/best.weights data/person.jpg
11、视频测试：./darknet detector demo data/voc.data yolov4-tiny-custom_2008.cfg backup/best.weights -thresh 0.25 data/1.mp4 -out_filename xxx.mp4
