last year I started an object detection project. Helpful answers from this community led me to the fastai 2018 course that helped me understand the problem. After quite some delays I want to give it another shot
My reimplementation only works to some extend, such that I’d like to use a library/framework to get rid of potential bugs in my implementation.
However, it seems that object detection compared to other problems isn’t that popular. As a consequence most repositories / tutorials I find are abandoned or outdated.
I just don’t find much information about how to finetune (or do transfer learning) with RetniaNet/SSD/YOLO.
Does anyone have recommendations for finetuning tutorials/support with pytorch / fastai?
Well, I would say that OD is popular but not very well supported within the fastai context.
Most libraries are kind of closed-shop so one has to pick a model along with its proprietary framework.
An exception is perhaps OpenMMLab’s MMdetection ( GitHub - open-mmlab/mmdetection: OpenMMLab Detection Toolbox and Benchmark) but that is not integrated with Fastai afaik.
There was the Icevision framework, too, but that is outdated. Some people are currently trying to revive it but I don’t know the status here. (GitHub - airctic/icevision: An Agnostic Computer Vision Framework - Pluggable to any Training Library: Fastai, Pytorch-Lightning with more to come)
Other opinions welcome though…
Object detection targets aren’t as consistently shaped as most other vision problems, so support is often lacking and people tend to use dedicated libraries for it.
detectron2 are both popular choices in the pytorch ecosystem.
If you are looking for something lightweight, without having to learn a framework such as Detectron or MMDetection there are some reference examples I have put together that may be useful, and should be easy to adapt: