/EFPN-detectron2

Primary LanguagePythonApache License 2.0Apache-2.0

EFPN implementation based on detectron2

To train:

run training scripts such as frcnn_x101_efpn.py as included here, change configs accordingly

Modifications from detectron2:

  1. "detectron2" folder name changed to "detectron" to correctly build (could also work outside of root dir)

  2. in detectron/modeling/backbone, we modified resnet.py and fpn.py to create the EFPN and FTT module

  3. slight modifications to fix import paths, cuda compatibility...etc, including in detectron/engine/defaults.py, detectron/layers/wrappers.py

Todo:

  1. create loss function for supervised and independent training of EFPN+FTT

Credits:

EFPN original paper: C. Deng, M. Wang, L. Liu, and Y. Liu. Extended feature pyramid network for small objectdetection.CVPR, 2020.

(https://arxiv.org/pdf/2003.07021v1.pdf)

Detectron2: Yuxin Wu, Alexander Kirillov, Francisco Massa and Wan-Yen Lo, & Ross Girshick. (2019). Detectron2. https://github.com/facebookresearch/detectron2.