A Python3.5/Pytroch implementation of RetinaNet: Focal Loss for Dense Object Detection. And the official implementations are available here. Besides, special thanks for those two repositories：

• pytorch-retinanet
• pytorch-retinanet

• python 3.5.x
• pytorch 0.4.1
• tensorboardX
• pillow
• scipy
• numpy
• matplotlib
• easydict

• Backbone: ResNet50

The training loss and accuracy :

VOC2007 detection results:

LUNA16: lung nodules detection. The mAP for different anchor scale and aspect ratios:

LUNA16 detection results:

• RetinaNet: neural networks and components that form parts of RetinaNet.
• config: define configuration information of Faster RCNN.
• data: scripts for creating, downloading, organizing datasets.
• loss: implementation of focal loss.
• pretrained_model: get and store pretrained ResNet model.
• targets: generate anchors and calculate targets.
• utils: tools package, containing some necessary functions.

1. Clone this repository (RetinaNet_pytorch):

    git clone --recursive https://github.com/Jacqueline121/RetinaNet_pytorch.git
   

2. Install dependencies:

    cd RetinaNet_pytorch 
    pip install -r requirements.txt
   

For PASCAL VOC, you can follow the instructions in this repository to download the data. And then, you can store date according the following structure:

|+-- data    
|   |+-- dataset    
|       |+-- VOC2007    
|           |+-- Annotations    
|               |+-- xxxx.xml    
|           |+-- Cache    
|           |+-- ImageSets    
|           |+-- JPEGImages    
|           |+-- Results    
|       |+-- VOC2012    
|           |+-- Annotations    
|               |+-- xxxx.xml    
|           |+-- Cache    
|           |+-- ImageSets    
|           |+-- JPEGImages    
|           |+-- Results    

• Annotations: store annotaion information(.xml file) for each images.
• Cache: store annotaion cache.
• ImageSets: store training dataset and testing dataset(.txt file) with the format:
• JPEGImages: store images.
• Results: store detection results.

You can also use your own dataset as long as you follow the file structure desribed above to store the data.

1. Download the pretrained ResNet model: ResNet50, ResNet101.

2. Put the pretrained model in $PROJECT/pretrained_model

3. cd $PROJECT/pretrained_model

    python get_pretrained_model.py
   

   It will produce a 'model.pth' file.

python train.py --dataset='Dataset_Name'

For example: python train.py --dataset='VOC2007'

python test.py --dataset='VOC2007'

If you want to visualize the detection result, you can use:

python test.py --vis