Where to start object detection with fastaiv2?

Hey everyone, how can I perform object detection with fastaiv2, I have some resources but they are built on fastaiv1, and it is not supported databunch of fastaiv2.
Thanks, everyone.

Hi there!

Right now, the most common tip you will receive is to use ice vision in combination with some fastai tricks.

This is one of the links that might be a good orientation point to start with.
For fastai object detection, you might have to listen to the new course or wait a while longer til the first results pop up here.


Does the new course have github repo where object detection is used?

As far as I know, this is currently under development. You could ask @jeremy whether there is a GitHub or whether object detection will even be a topic.

On the other hand, @muellerzr has also a revisited course that includes segmentation and parts of objection detection.
Since I stumbled yesterday upon a link to his uncommented notebooks here, I think he would be okay if they are shared as well

But he is doing a course as well and you might want to consider getting any or both of them. Both are really good teachers and their content is very good and contains always something new or funny.


IceVision is the standard recommendation for object detection in-conjunction with fastai.

Developers and contributors are members of this community.


icevision doesnt support medical dataset files like WSI microscopic large images with tiff format and MS COCO annotations, till now i haven’t found a single notebook or resource using that.

this doesn’t have object detection , only semantic segmentation and other basics of fasai nbdev and fastcore.

Yes indeed, I open sourced them for that reason :hugs: (This community ofc has a special place in my heart, so here I’m okay with them being posted :slight_smile: )

Re; object detection, while it doesn’t, it teaches you how to step away from the fastai framework for many things, so that you can look and approach PyTorch code and existing implementations and not be shackled if fastai can’t do something “perfect”. If you think that’d be useful for you, feel free to get the course (its on sale rn). If not that is OK :slight_smile: (as I personally don’t recommend fastai OOTB for O.D.)

There was no mention of medical imagery in your question.

@farid do you have any examples on medical data formats?

1 Like

@johnrobinsn has recently published a series on object detection with fastai2 which seems pretty great.


Here’s his articles, cool!! :slight_smile:


yeah that is the main problem, I have object detection examples with pytorch also, but loading the medical data is where the problem is , I have used a library that uses slide-container to load the data : GitHub - ChristianMarzahl/ObjectDetection: Some experiments with object detection in PyTorch
But the problem here is that it is based on fastaiv1 , so it creates dynamic databunch object,
that is there are changing batches for my data.
I wanted a data bunch where the batches of images are fixed.
How can I achieve this?

I just posted the final article in my series on SSD Object Detectors.

Here is my tweet if you’d like to give me a like :slight_smile: