GraphFPN: Graph Feature Pyramid Network for Object Detection with Retinanet

Inspired by Graph_FPN and RetinaNet, we have used the Graph_FPN structure as a backbone to train on Retinanet and the work is not yet complete.

Currently working with the author Gangming Zhao, we will improve our code on mmdetection in the future.

For demonstraction with Graph_fpn, run:

python demo.py

For demonstraction with fpn, run:

python demo.py --no_graph

For training , run:

python train.py

For test with Graph_fpn, run

python test.py

For test with fpn, run

python test.py --no_graph

If You need COCO API for test, you can download from here. You need to set the backend of DGL to tensorflow, here is tutorial link

Folder structure

${ROOT}
└── checkpoint/
└── COCO/    
│   └── coco/
│   │    ├── .config 
│   │    ├── 2017/
│   │
│   ├── downloads/
│
│
└── data_demo/
|   ├── data/
|   |    ├── coco
|   |    ├── checkpoint
|   ├── data.zip
|
├── results/
├── src/     
|   ├── configs/
|   |    ├── configs.py
|   |
|   ├── detection/
|   |    ├── datasets/
|   |    |      ├── coco.py
|   |    ├── utils/
|   |
|   ├── model/
|   ├── init_path.py
|   ├── demo.py
|   ├── train.py
|   ├── test.py
├── README.md 
└── requirements.txt

References

[1] Retinanet: Focal Loss for Dense Object Detection
[2] Graph-FPN: GraphFPN: Graph Feature Pyramid Network for Object Detection
[3] Object Detection with RetinaNet: Keras Implementation