Fastai_object_detection

fastai
1

Hi all,

I created a extension for fastai to include object detection. Currently there are models like FasterRCNN, MaskRCNN and EfficientDet included. It comes with a class, which makes it easy to download images and annotations from MS COCO to create your own custom object detection/ instance segmentation dataset. Try it out, maybe you find it useful.

The code is public on github.
The extension is build with nb_dev, so it’s build in jupyter notebooks and documentated.
You can find the documentation here.

10 Likes
2

Wow. great work. Thanks a lot

3

This is awesome. Any plan to add YOLO5?

6

that is excellent news, and very interested in yolo5 as well

7

Unfortunately YOLO5 is under GPL license, so I can’t use the code and publish it in my Apache-2.0 licensed repo…

But there is a guide how you can implement your own models here.

Maybe I will create a second repo with just the yolo4 or 5 model for the next kaggle competion, when I have time for it. :smiley:

2 Likes
8

Thanks for this Robert!

I’m working with your extension and have found it quite useful to get a Faster RCNN model running quickly! Now, I’d like to use the ColSplitter() instead of RandomSplitter() with the ObjectDetectionDataLoaders class. This seems like a straightforward modification to your source code, though I’m having some difficulty in the implementation. Any help would be really appreciated!

9

Hello,

Thank you very much for creating the extension, it is really great! Could you recommend the best way to create DataLoader from a custom dataset (images and annotations) and not COCO dataset?