zappytoes / nightingale

Nightingale: State-of-the-art object detection with object-aligned-boxes

Welcome to Nightingale v0.1!

Nightingale is a Python & Tensorflow codebase for training, exporting, inferencing, & evaluating the state-of-the-art SCRDet++ object detection model with ResNet backbone on NITF imagery. Nightingale's inference module is compatible with the OMITTED suite of products. The current version of Nightingale is configured for system single channel imagery.

Some exciting features of Nightingale include:

• A pre-configured Tensorflow-GPU anaconda environment or Docker container

• State-of-the-art detector performance for Object-Aligned Bounding-Boxes (OAB) with the SCRDet++ model

• Easy-to-follow model training & inference guides with Jupyter Notebooks

• An updated training pipeline for NITFs, with background-only (i.e., annotation-free) image support for enhanced False Positive reduction

• A module for quickly exporting trained models to a deployable Tensorflow frozen-graph format

• An intuitive inference API with OMITTED-compatible I/O

• A fast & powerful inference module for processing large NITFs using a parallel-gpu image-tiling scheme

Who should use Nightingale?

Nightingale is for Computer Vision data scientists, researchers, or developers with access to the Intelligence Community (IC) Common Operating Environment (COE). You'll need a linux style GPU-accelerated computing environemnt with Jupyter Notebooks, such as users with a Redmine RLE account and access to RLE's JupyterHub or DGX systems. It is also for those interestd in object-aligned object detection or a OMITTED-compatible detector.

More info:

Links omitted

Getting started

You'll need access to a linux style operating system with Jupyter Software (or similar IPython notebook interpreter) installed.

Detailed instructions for getting set up and executing Nightingale with code and data examples are detailed in a series of Jupyter Notebooks contained in this repo. Open the "Welcome_To_Nightingale.ipynb" to get started.