Questions about small object detection

[HoracceFeng]

Hi @avanetten ,

Thanks for your work. I have read some articles talking about your paper and since now I am working on traffic-sign detection (though it is a much easier problems than yours), it will be great if we can make some discussion on this topic.

From my perspective, your model generally based on YOLO2, and the improvement are mainly on:

1. crop the big picture to several small one, and than detect each
2. modify the network so the last feature map will be 26^2 but not 13^2
   Am I right?
   Actually, I am now using YOLO3 to detect traffic sign. The picture in my dataset is 2048^2, the size of ground truth boxes vary from 10 pixels to around 100 pixels. Like you, I also do random crop to make the input size become 512. Unfortunately, the result is pretty bad. I have also tried other input size such as 1024, but the result seems similar.

So, have you tried using your strategies in YOLO3? And, any other structures or parameters modification do you think will do good to small object detection?

Many thanks.