search

facebookresearch / detectron2

Detectron2 is a platform for object detection, segmentation and other visual recognition tasks.
https://detectron2.readthedocs.io/en/latest/
Apache License 2.0
30.64k stars 7.5k forks source link

vgg16 backbone #1209

Open chansongoal opened 4 years ago

chansongoal commented 4 years ago

🚀 Feature

VGG16 is a universal network and have been used as a backbone in FasterRCNN. Is there a VGG16-based model in Mask-RCNN?

ashnair1 commented 4 years ago

Hi @ppwwyyxx , I'm interested in writing up a PR for the VGG models but I had a couple of questions before I proceed.

  1. I'm thinking the VGG model should have a separate base yaml i.e. Base-RCNN-VGG.yaml so it can be later be sub-classed to specify different depths (16, 19 etc). Do you agree?

  2. There's also the issue of incorporating pretrained vgg models. Should something similar to torchvision_to_d2 script be done?

Alternatively, I could just submit a PR and we can continue the discussion from there as these questions are implementation specific.

reactivetype commented 3 years ago

Is there an update for merging the PR?

ashnair1 commented 3 years ago

detectron2 has gone through a lot of changes since I last worked on this PR (~9 months ago), so it most likely isn't compatible. I can work on it if there is interest. Feedback on the PR would be appreciated so that I know I'm on the right track.