Introduction

KWIVER is a fully featured toolkit for developing Computer Vision Systems, a capability that goes beyond the support of simply developing Computer Vision Software.

This distinction is an important one. There are myriad of software frameworks that facilitate the development of computer vision software, most notably the venerable OpenCV, but also including VXL, scikit-image and a wide range of others. The current Deep Learning revolution has additionally spawned a number of software frameworks for doing deep learning based computer vision including Caffe, PyTorch, Tensorflow and others.

Each of these frameworks has their own unique set of capabilities, target user community, dependencies and levels of difficulty and complexity. When developing computer vision software, the task frequently boils down to selecting the most appropriate framework to work with and proceeding from there.

As the task at hand becomes more complicated, however, the burden on the supporting frameworks, and the task-specific software developed using those frameworks, becomes heavier. Real world problems might be better solved by, for example, fusing OpenCV based motion detections with Faster-RCNN (Caffe) based appearance detections and then filtering the result against a new state-of-the-art image segmentation neural network that runs in yet another deep learning framework. Couple this with the understanding that computer vision algorithms traditionally are extremely compute intensive, doubly so when one considers the GPU requirements of modern deep learning frameworks and it is clear that building computer vision systems is a daunting task.

KWIVER is designed and engineered from the ground up to support the development of systems of this nature. It has first class features that are designed to allow the development of fully elaborated systems using a wide variety of computer vision frameworks – both traditional and deep learning based – and a wide variety of stream processing and multi-processing topologies. KWIVER based systems have scaled from small embedded computing platforms such as the NVIDIA TX2 to large cloud based infrastructure and a wide variety of platforms in between.

KWIVER is a collection of C++ libraries with C and Python bindings and uses an permissive BSD License.

Visit the repository on how to get and build the KWIVER code base