Problem
A two-person team (one developer, one data scientist) at RocketControl builds computer vision for mining plants. The first case is froth analysis on flotation machines at an on-premises plant. The servers run Astra Linux with no GPU, and the business expects 3 or more plants. Without a standard pipeline, every task keeps both people at 100% load and takes several weeks, mostly because of annotation.
Example from the article: in production the cluster of 3 servers (CPUs from 2017, no GPU) ran 38 cameras with the same number of applications at an average system load of about 220%, and the applications froze.
Solution in brief
Three libraries wrap the solution, and a toolbox lets process engineers do annotation, frame preparation and validation themselves. Deployment goes through GitLab CI/CD. After scaling, the team fixed frame capture and CPU load.
Questions this case answers
- How to standardize a computer vision pipeline so a small team can roll it out to many sites with minimal adjustment and move annotation and validation to non-technical staff
- How to run video-based detection on CPU-only on-premises hardware when dozens of concurrent streams overload the system
- How to standardize a computer vision solution so it can be rolled out to many factories with minimal adjustment
- How to cut data scientist and developer load in a CV project by giving process engineers annotation and validation tools
Tools and methods
- cv-toolbox — tools for engineers: video annotation, frame tasks, polygon zones, detector and tracker validation, outsourced annotation
- cv-mdk and cv-product — model creation, frame processing, and the solution wrapper
- CVAT with Jira — annotation tasks, synchronized through both APIs
- SAHI and SAM — semi-synthetic pre-annotation of frames
- OpenMMLab (RTMDet with BoT-SORT), OpenVINO, TensorBoard — detection, tracking, CPU inference, monitoring
- GitLab CI/CD — packaging and deployment to production instances
- GStreamer — frame capture instead of ffmpeg
Google or a one-time email code. No password.