Instance Segmentation / Object Detection

I’m curious if there has been any update to the ability for Instance Segmentation in fastai over the past two years? An example in FastBook that uses [ImageBlock, BBoxBlock, BBoxLblBlock] is conspicuously missing.

The most I’ve seen is a few outdated notebooks that no longer run (by muellerzr, and sgugger) dispersed across github. From what I can tell, they seem to require a ton of labor to reimplement RetinaNet (is there a newer better solution?), and even then, show_results() and predict() don’t work according to “A walk with fastai2”.

Is there a plan in the works, or is the idea to just let 3rd parties like IceVision handle object detection?



There is no proper object detection support in fastai v2 unfortunately. Zach’s notebooks should still mostly work though, did you get an error?

Additionally, I think @KevinB might have explored object detection with fastai v2.

1 Like


The best for Object detection is to use the Yolo family, Yolov5, YoloR, and YoloX, Yolov5 has won several competitions on kaggle; from my point of view, it is like an AutoML for object detection, you set the parameters, and the framework does everything else for you. Of course, there is cleaning up on the images, etc, etc., etc.

1 Like

I’ve been waiting for the object detection world to settle on a more reasonable approach – I plan to revisit it in part 2 this year.


Thanks for the update–looking forward to it!

I have been working a lot with Obj Detection, and I will love to contribute if possible to FastAIObj.