Yolov3_Tiny_VisDrone

Training yolov3_tiny on VisDrone 2019 detection dataset

Platform

GPU : NVIDIA Quadro RTX 6000/8000
CUDA toolkit version : 11.4
Driver version : 470.57.02
cuDNN version : 8.8.1.3
OpenCV version : 4.2.0
Ubuntu : 20.04

Resources

Dataset : https://github.com/VisDrone/VisDrone-Dataset
Darknet installation guide : https://github.com/ERYAGNIK003/darknet
Download convolution layers(darknet53) weights : https://pjreddie.com/media/files/darknet53.conv.74
Transform VisDrone data format to Darknet : https://github.com/zhaobaiyu/visdrone
Training instruction : AlexeyAB/darknet#504
visualizing network : https://github.com/hahnyuan/darknet-visualizer

Directory structure

|
|___darknet
|
|___VisDrone
         |__VisDrone2019-DET-train
         |__VisDrone2019-DET-val
         |__VisDrone2019-DET-test-dev
         |__dataTransform.py

Steps

• Setup CUDA driver, cuDNN library, OpenCV and Darknet on your system as per given in https://github.com/ERYAGNIK003/darknet
  
• Download darknet53 weights into darknet directory
  • cd darknet
  • wget https://pjreddie.com/media/files/darknet53.conv.74
• Download VisDrone 2019 DET dataset and organise darknet and visdrone as per given directory structure
• Download dataTransform.py from repository and store into VisDrone directory, this will helps to convert data format
• Execute it by "python dataTransform.py" , this will create images.txt file in all three dataset directory
• Add configuration files into darknet directory
  • Download visdrone.data into darknet/cfg/
    • Modify "path-to-dataset/images.txt" according to your dataset path
  • Download visdrone.names into darknet/data/
  • Download yolov3_visdrone.cfg into darknet/cfg/
• To visualize your netowrk use graph visualizer
  • pip install graphviz
  • sudo apt-get install graphviz
  • git clone https://github.com/hahnyuan/darknet-visualizer.git
  • cd darknet-visualizer
  • python3 darknet_visualize.py example/yolov3.cfg (replace path to .cfg file as per yours)
• Start training yolov3_tiny on visdrone dataset using below command,
  • cd darknet
  • ./darknet detector train cfg/visdrone.data cfg/yolov3_tiny.cfg darknet53.conv.74 -dont_show -map

Testing

• ./darknet detect cfg/yolov3-tiny.cfg backup/yolov3-tiny.weights data/1.jpg