Tackling the Small Object Problem in Object Detection

Small objects are hard to detect!
In this video, we discuss some strategies to solve the small object problem in object detection.
Corresponding blog post:
blog.roboflow.com/tackling-th...
✅ Subscribe: bit.ly/rf-yt-sub
💻 Project Templates: roboflow.com/templates

[@MustafaAktasVEVO]

Helpful video, thank you.
Another way to augment data might be to go and get a high-resolution dataset and decrease its resolution so that it will be like small objects.

[@ekzacharia]

This was the most helpful video I've watched all year!

[@ekzacharia]

Just like so many questions answered.

[@TheTacticalDood]

I didn't get why we should filter null tiles. Shouldn't they help teach the model the "absence" of objects?

[@NadavOssendryver]

Can you please help me, how to I add tiling at inference, and not only in the preprocessing?

[@fedorlaputin9119]

how did you solve that?

[@wittman99]

@@fedorlaputin9119looking at documentation you will need to pass &tile parameter and set to the size of tile which you used but this option does nothing and run inference on the whole image. Im trying to figure this out too.

[@zoroXgamings]

I have 200mb of data as tfrecord file, while training efficientdet d7 , it stops

[@gjergjishkurti4061]

check that you're not running out of memory