tiantuxu/darknet

YOLOv3 & YOLOv2 that is optimized for running object detection on consecutive video frames

YOLOv2 & YOLOv3 for Consecutive Video Streams

Requiremets:

• Linux GCC>=4.9
• CUDA >= 7.5
• CuDNN >= 5.1
• OpenCV 3.3.0
• ffmpeg (with NVIDIA Support): https://developer.nvidia.com/ffmpeg

Build

$ export PATH=/usr/local/cuda-8.0/bin${PATH:+:${PATH}} 
$ export LD_LIBRARY_PATH=/usr/local/cuda8.0/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
$ make

Get YOLO weights

Download YOLOv2 or YOLOv3 weights

Video Preparation

Preparing Videos and decode to images: Transcode to 608x608 & Take out 10 min video sample (Optional)

$ ffmpeg -c:v h264_cuvid -i input.mp4 -ss 00:00:00 -t 00:10:00 -vf scale=608:608 -c:v h264_nvenc output.mp4

Image Generation

Decode and save as images, with filenames defined in frame sequences

$ ffmpeg -i input.mp4 ./path/to/images/%7d.jpg 

Generate filename lists, and the text file will be in ./data/train.txt; Remember to change path to images in generate-name.py

$ python genearte-name.py

Run it!

Note: Can change to all availble cfg under./cfg

$ ./darknet detector test cfg/imagenet22k.dataset cfg/yolov3.cfg yolov3.weights -dont_show -ext_output < ./images.txt > result.txt

Inference results is in result.txt

/path/to/image/0000000.jpg,OB1,97% (left_x:  xxx   top_y:  xxx   width:  xxx   height:  xxx)
/path/to/image/0000000.jpg,OB2,90% (left_x:  xxx   top_y:  xxx   width:  xxx   height:  xxx)
/path/to/image/0000001.jpg,OB2,88% (left_x:  xxx   top_y:  xxx   width:  xxx   height:  xxx)
/path/to/image/0000003.jpg,OB3,96% (left_x:  xxx   top_y:  xxx   width:  xxx   height:  xxx)
/path/to/image/0000005.jpg,OB4,89% (left_x:  xxx   top_y:  xxx   width:  xxx   height:  xxx)
...