This repo aims to implement a "Focal loss for Dense Object Detection"
- python 3
- Pytorch 0.4
- torchvision
- numpy
- tensorflow
# Select the script that you want to train for reproducing a results
./retina_ce_sgd_0.001.sh
# For the focal loss use ./retina_focal_sgd_0.0001.sh
You can see the details in trainer.py
#####Download VOC2007 trainval & test
# specify a directory for dataset to be downloaded into, else default is ~/data/
sh data/scripts/VOC2007.sh # <directory>
#####Download VOC2012 trainval
# specify a directory for dataset to be downloaded into, else default is ~/data/
sh data/scripts/VOC2012.sh # <directory>
- Building RetinaNet
- Apply cross entropy loss and focal loss(still checking it's works or not)
- Comparing btw CE loss and focal loss
- Report results on VOC
- currently i get 50mAP on VOC2007. There is some problem that i don't know yet.
- Report results on COCO
- Use relative path for easy reproducing of result
- Keras version
- Easy to read and very active repository. This repo drive me complete my code
- Another pytorch version
- Almost all code are followed from this repository, but i am not sure this repo works
jtlee at omnious dot com
Feel free to contact me about code.