search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.78k stars 7.96k forks source link

External Data augmentation vs Training time data augmentation #7922

Open snake-arch opened 3 years ago

snake-arch commented 3 years ago

I am trying to detect small objects and have come across tiling. But now I wonder what is the difference between internal and external data augmentation. Can external bring positive changes? If so why don't we use it often?

stephanecharette commented 3 years ago

I use YOLOv4 to find objects down to 7x7 pixels in images measuring 3840x2160. I have youtube tutorials showing how exactly I do it. For example, see https://www.youtube.com/watch?v=gPP6fh8IIAo&t=174s

This is done using DarkMark and DarkHelp in combination with Darknet and YOLO. You can find more videos and explanations in my youtube channel: https://www.youtube.com/c/StephaneCharette/videos

As to your question "why don't we use it often"...? I use it every day.

Nuzhny007 commented 3 years ago

@stephanecharette and what about net input size? Do you scale all frame or cut it into crops? And it works in real time? Thnx!

stephanecharette commented 3 years ago

@Nuzhny007 see this video for example: https://www.youtube.com/watch?v=7e0Gckun1Ds

Nuzhny007 commented 3 years ago

Thank you!