What's the next goal or roadmap 2022 in Detectron2?

[p890040]

🚀 Feature

Detectron2 is amazing repo for instance segementation open source and toolbox.

Even it only focus on mask-rcnn architecture, the scalability and flexibility in detectron2 are still strong.
However, there is no update since last year.
I'm a big fan for Detectron2. I'm looking to new feature or any progress.
Is there any roadmap or plan information?

[Kevin-Delnoije]

I am wondering this too, I think this repo is still amazing but the model zoo is lacking a bit.
Here are some thoughts of things to improve.

detection models

• improved model architectures based on FCOS, (e.g. VFNET, gfocalv2, TOOD)
• newer backbones with pretrained final model (efficientnetv2/swinv2)
  • efficientnetv2 is currently in torchvision master
  • swinv2 code when it is released or maybe wait for transformer evolution to stabilize
• stronger neck (bifpn)

All the above things can easily be done inside projects outside the main repository, however the main problem I have
with current projects is they get outdated and unmaintained. It would be nice to have a project with a few sota like architectures and their weights pretrained.

The blog post on improved augmentation/schedule for mask-rcnn is really nice, but is only mask-rcnn.
These weights are nontrivial to use for object detection/ keypoint detection and other architectures like
retinanet/fcos would also benefit from this approach.

semseg models (encoder decoder like)

One frustration I had was changing the head to something that was not directly pixel classification but instead pixel regression. I found it nontrivial to do this while in theory it is only changing head and loss of the architecture. a tutorial notebook/project
applying this on a toy example would be a great addition.

more tutorials/examples

documentation and overview tutorials are great, but people getting started often want some copy/paste examples for inside a notebook to play around with some tutorials that are specific on one topic can be a nice addition.
topics can be:

• finetuning
• modifying small part of network, debug/iteration method
  • in jupyter notebook or script or config and iterate over design
• custom evaluation
• custom data augmentation
• custom structures

I am happy to contribute any of the above if it aligns with the roadmap

[mahilaMoghadami]

where can I download swinv pretrained model?