Train Detector on MS COCO (trainvalno5k 2014) dataset

Train Detector on MS COCO (trainvalno5k 2014) dataset

Note: For using Mosaic data augmentation and to avoid the bottleneck of CPU performance for data augmentation, compile Darknet with OpenCV.

Required files:

• yolov4.cfg - https://raw.githubusercontent.com/AlexeyAB/darknet/master/cfg/yolov4.cfg

  (change width=512 height=512 in cfg-file)

  mini_batch_size = batch / subdivisions, so higher subdivisions= - lower accuracy (AP):

  • for 32 GB GPU-VRAM set subdivisions=8 in cfg-file (yolov4.weights is trained on V100-32GB)

  • for 16-24 GPU-VRAM set subdivisions=16 in cfg-file

  • for 8-12 GB GPU-VRAM set subdivisions=32 in cfg-file (if Out Of Memory occurs - set random=1.34 for [yolo] layers)

  • if you compiled Darknet without OpenCV and disabled mosaic=0 then accuracy may be ~2% lower

• csdarknet53-omega.conv.105 - The first 105 layers from the pre-trained weights on ImageNet (ILSVRC2012):

  • download weights csdarknet53-omega_final.weights for Classifier: https://drive.google.com/open?id=18jCwaL4SJ-jOvXrZNGHJ5yz44g9zi8Hm
  • get only the first 105 layers ./darknet partial cfg/csdarknet53-omega.cfg csdarknet53-omega_final.weights csdarknet53-omega.conv.105 105

• data/coco.names - with names of objects https://raw.githubusercontent.com/AlexeyAB/darknet/master/data/coco.names

• bash-script to get MS COCO training dataset and unzip archives (20 GB): https://raw.githubusercontent.com/AlexeyAB/darknet/master/scripts/get_coco_dataset.sh

  • trainvalno5k.txt - paths to Training images
  • train2014.zip and val2014.zip - Training images (trainvalno5k 2014)

• cfg/coco.data - with content (testdev2017.txt isn't required for training):

classes= 80
train  = <replace with your path>/trainvalno5k.txt
valid = <replace with your path>/testdev2017.txt
names = data/coco.names
backup = backup
eval=coco

---

Training

Note: The results in the paper were obtained using training on 1 GPU (all parameters in yolov4.cfg by default, except for width=512 height=512)

• Training command:

./darknet detector train cfg/coco.data cfg/yolov4.cfg csdarknet53-omega.conv.105

• Continue training:

./darknet detector train cfg/coco.data cfg/yolov4.cfg backup/yolov4_last.weights

You will get files

• chart.png with Loss chart
• backup/yolov4_last.weights - trained last weights file
• backup/yolov4_final.weights - trained final weights file (when training is complete)

---

Detection:

Detect on the image:

./darknet detector test cfg/coco.data cfg/yolov4.cfg backup/yolov4_final.weights dog.jpg -ext_output

Detect on the video-file:

./darknet detector demo cfg/coco.data cfg/yolov4.cfg backup/yolov4_final.weights test.mp4 -ext_output

Detect on the video-file by using GPU-1:

./darknet detector demo cfg/coco.data cfg/yolov4.cfg backup/yolov4_final.weights test.mp4 -ext_output -i 1

Detect on the RTSP/HTTP video-camera:

./darknet detector demo cfg/coco.data cfg/yolov4.cfg backup/yolov4_final.weights http://192.168.0.80:8080/video?dummy=param.mjpg -ext_output