Facebook AI Research (FAIR) has announced the release of Detectron2 – a PyTorch-based object detection library as the second version of Detectron, released last year.
The new library is built from scratch, therefore, it’s not an update of Detectron but a complete rewrite in a more modular way, according to engineers from FAIR. Detectron2 is implemented on top of PyTorch and has numerous improvements over the previous version.
Engineers from Facebook AI Research mention that Detectron2 was built with the purpose of rapidly moving research ideas in the context of object detection into production models that can be deployed at scale. According to them, the new object detection library will help researchers and practitioners to quickly develop prototypes and iterate more rapidly on model design and experiments. They say that Detectron2 has a flexible and modular design that will allow users to plug custom module implementations for different parts of an object detection system.
The new library comes with many new features as well as many new models which can be easily accessed for usage. It includes all the models from the previous version of Detectron like Fast-RCNN, Mask R-CNN, RetinaNet, and DensePose, and new models like Cascade R-CNN, Panoptic FPN, TensorMask, etc.
One of the important novelties in Detectron2 is the distributed training that can be conducted over multiple GPU servers in an easy way, and also the completely GPU-based training pipeline. Moreover, engineers from FAIR created Detectron2go which is an additional layer that will allow easy and optimized model deployment to cloud and mobile.