tcalexcoassin/cvdetect repository overview

A pair of images, one a base to streamline builds and one containing the code to build out a computer vision system for processing RTSP streams from cameras. Currently, it is compiled to support CUDA 10.2 with Pascal generation NVidia cards.

The base image is built from the CUDA 10.2 Runtime image for Ubuntu 18.04 with CUDNN 7, Python 3.6 and OpenCV 4.3.0 compiled with CUDA support.

Running these images requires that the NVidia runtime for Docker be installed and used with these images and should allow for GPU accelerated use of OpenCV libraries within a Docker container.

The application image contains a configurable Python application which can use OpenCV to read an RTSP stream and detect objects using some pre-trained neural nets and output detection images and videos as well as periodic snapshots.

Unfortunately, even the NVidia Runtime image is quite large, and combined with the dependencies pushes the image size up to almost 3GB. But, this is much better than starting from some of the 6-8GB Tensorflow images. I hope to provide the Dockerfile and code on Github eventually for anyone else interested in building their own GPU accelerated OpenCV container, as I don't foresee myself attempting to produce images for every possible variant of NVidia card and driver version.