Yolo for Detectron2

Implementation of Yolo using Facebook's Detectron2 Framework.

With added quantization support following the work in

Chen, Peng, et al. "Aqd: Towards accurate quantized object detection." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021.

Description

This repo implements YoloV5 within Facebook's Detectron2 framework. Currently, only YoloV5m has been fully tested. Support is included for YoloV4-tiny. Support will be extended for other Yolo versions.

This repo also enables quantization and quantization-aware-training using the framework provided in

Chen, Peng, et al. "Aqd: Towards accurate quantized object detection." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021.

The quantization framework is implemented in QTool: A low-bit quantization toolbox for deep neural networks in computer vision. Use the quantization branch to train and test quantized models.

Setup

• This repo requires installation of the Detectron2 framework.
• Installation instructions can be found on their own GitHub page.
• However, if you intend on using the quantization modules, follow the instructions given in QTool - Detection to install the quantization fork of the detectron framework. Currently, Yolo has only been tested within this framework.
• Once Detectron2 has been installed, clone this repo. You can clone it anywhere you like, but it is convenient to clone it to the detectron2/projects directory for consistency.

  git clone https://github.com/ShechemKS/Yolo_Detectron2.git

  That's it! It will just work.

Training and Inference

To train the model run

python train_net.py --config-file configs/yolov5-Full.yaml

You may include any of the usual Detectron2 config options.

To use the model for inference, run

python inference_net.py --config-file configs/yolov5-Full.yaml --inputs path/to/image-dir/ --output path/to/save-dir/

TODO

• [x] Create a non-quantization branch for compatibility with the main detectron framework
• [ ] Add support for other Yolo versions

References

• facebookresearch/detectron2
• MonashAI/QTool