Find a new object detection backend (computer vision)

[raphael0202]

Currently, 3 object detection models are in production, served by Tensorflow Serving:

• nutriscore grade detector
• nutrition table detector
• universal logo detector

All these models were trained using Tensorflow object detection repository.. However, this repository is not actively maintained. It is also not compatible with Tensorflow 2.0 (due to the reliance on tf.contrib), which prevents us from using newer versions of Tensorflow Serving. Finally, more recent object recognition models/architectures are not available on Tensorflow Object Detection API.

Hence, we need to find a project that will allow us to train object detection models. Requirements:

• ease of training/deployment
• well maintained
• Inclusion of modern architecture/models
• Pytorch or Tensorflow. In both case, it should be servable through Tensorflow Serving or Pytorch Serve. If Tensorflow, should be compatible with Tensorflow 2.0.

To test new repositories, all train/val data and current models are available on Robotoff model repository.

Candidates:

• https://github.com/facebookresearch/detectron2
• https://github.com/zylo117/Yet-Another-EfficientDet-Pytorch