Detectron2 by Facebook AI Research is Here!

Learning In Chunks

“Detectron2” is latest release(PyTorch) of Facebook AI Research(FAIR)’s popular Object Detection Platform and it is a ground-up rewrite of earlier version (Caffe2). The platform comes with support for single or multiple GPU for fast training .

Source: facebook

New additions are:

• Implemented with PyTorch deep learning framework.
• Comes packed with Faster R-CNN, Mask R-CNN, RetinaNet and new Features such as State-Of-The-Art(SOTA) object detection algorithms like:- DensePose, Panoptic FPN, TensorMask^, Mesh R-CNN^ (^- releasing soon).
• Added features such as synchronous Batch Norm and support for new datasets like LVIS(https://www.lvisdataset.org/).
• Training is Faster now with multiple GPU support .
• Comparison Chart (Mask R-CNN in detectron2):

Source: facebook

SOTA Comparison Table (July 2019):

Source : Bing Search

Hardware used for testing Detectron2:

Source: facebook

System Setup:

Source :https://github.com/facebookresearch/detectron2/blob/master/INSTALL.md

Quick Start using Colab:

Link - https://colab.research.google.com/drive/16jcaJoc6bCFAQ96jDe2HwtXj7BMD_-m5

Results of Test Run:

Instance Segmentation (Mask-RCNN-R-FPN50)

Panoptic Segmentation

DensePose

Feels nice to see Big Tech giants like Facebook, Google, OpenAI, Microsoft, Nvidia are coming up with more and more opensource research which boosts and opens up the scope just not only from Object Detection but expands to Pose Detection, Instance & Semantic Segmentation , Self-Driving Vehicles and more.

Keep Learning !!!

if liked :-)