Hi all,
I’m looking for any pre processing technique that can help with detecting very small object in pictures, for example.
If I just put the annotation in, the results are very poor
Hey there!
It’s been a long time since this was originally posted; however, i thought I would add some resources I have found and maybe this kickstarts this topic again.
- Depending on what model you’re using for these detections, you can get the best anchor sizes for your RPN by refining them to the bounding boxes in your dataset. I have added more in the link here: What are Anchor Boxes in Object Detection?
- SAHI (Slicing Aided Hyper Inference): Essentially cuts the image into tons of different, smaller windows and then inferences over them. Paper: https://arxiv.org/pdf/2202.06934
That’s all i really have for the moment! If you happen to still be working on this or if anyone reading this has a similar issue, don’t be afraid to reach out!
1 Like