Project information

• Category: MLOps
• Client: INSTAR Deutschland GmbH (Waletech) Shenzhen, China
• Project date: 15 Jul, 2026
• Project URL: INSTAR Cloud

The problem: Our cameras ran a legacy Caffe detector — fast but low-confidence, able to tell only three generic classes apart. Customers kept filing the same two complaints: it flags things that aren't there, and it misses things that are

What I did: I turned "a better model" from a guess into a process I owned end to end. I benchmarked five architectures against real surveillance footage and chose YOLO for speed, precision and ease of use. Then I went to the customers and pulled their complaint videos — that feedback drove the dataset, not the other way around.

I built a surveillance-specific dataset for what these cameras actually see: grainy frames, motion blur, low light, and monochrome infrared-LED night video. And extended the model from three classes to nineteen, because that's how many things customers said they needed to see.

From YOLOv3 to YOLOv10 I picked YOLOv7 for embedded edge deployment — the best balance of precision, size and speed for the target hardware. And YOLO10 for the cloud deployment for it's higher accuracy and ease of export to CPU-friendly formats (float16 / RTLite). I consolidated the embedded Linux and cloud teams' needs so one model type served both, simplifying the training steps, then exported the PyTorch model to ONNX, quantized it, and converted it to the proprietary Novatek firmware format.

Once deployed I fed follow-on complaint videos from live customers into an MLflow-monitored training pipeline. Accuracy climbed and false detection dropped on both sides — fewer false positives and fewer false negatives. After that model held in the field I built a second one around static objects: parking, missing protective gear, fire and smoke, unattended luggage

The payoff: The same training pipeline for all models, evaluated on real footage, served as a canary-deployed Nomad job for the cloud with Zabbix dashboard and added to a Gitlab CI/CD pipeline for the camera firmware build.