AI/ML: video stream tagging

Create a system to upload user videos for further analysis. The system should be capable of analysing the video automatically. The system must provide for 10 concurrent user sessions and be linearly scalable to increase the number of simultaneous users up to 500.

To solve the problem, we used state-of-the-art machine learning models. The models are based on convolutional neural networks, which allows us to speed up the learning and inference tasks by using GPU. Thus, if there are sufficiently powerful GPUs in the cluster, it is possible to achieve near real-time prediction performance (the number of frames per second processed by the model approaches the video FPS value). The module allows to mark up video (photo) datasets in parallel, which is later used for model training.

We developed a multi-node linearly scalable platform for automatic video analysis using GPUs, which made it possible to significantly increase the performance per cluster node. A user interface was developed for maintaining a video library, viewing and analyzing it, as well as allowing you to upload videos and select ML models to be used for analysis.