Small Object Multiscale Real-time?

[whittenator]

I am in need of a SOTA Object Detection model framework that I can train a custom dataset with. I have trained Yolov7 on the dataset but it seems to still struggle with the varying scale of objects in the dataset. The dataset has objects ranging from 5x5 pixels to nearly taking up the entire 1920x1080 screen. Can anyone recommend a model framework for working best in this situation? This model needs to be able to run in real-time at 30FPS on Nvidia devices(Jetson or Nvidia Server).

Thanks

[hankyul2]

Hi @whittenator

How about using YOLO-MS?

• They achieve a better trade-off between accuracy and computation cost.
  
• They capture small-scale objects as well.
  

I hope this could help you.

Thank you.

Hankyul.