SSD with FPN (RetinaNet) in tensorflow

Backbone	AP	[email protected]	inference time
Mobilenet v1	31.7	49.8	36 ms
Shufflenet v2	29.3	46.7	31 ms

Notes

1. Inference time was measured on images of size 896x640 and using NVIDIA GTX 1080 Ti.
2. Average precision was computed on COCO val 2017 dataset.
3. During the evaluation I resize smallest dimension to be equal to 640 (while keeping the aspect ratio).
4. You can get the pretrained models from here.
   You can test them using inference/just_try_detector.ipynb.

How to use this

1. Prepare dataset. See data/README.md.
2. Edit a json configuration file and then just run python train.py.
3. Then use create_pb.py to get a frozen inference graph.
4. Use notebooks in inference folder to test the trained detector.

Requirements

• tensorflow 1.12
• tqdm, Pillow, numpy, matplotlib

Credit

This implementation is based on Tensorflow Object Detection API.