This repo is an extension of the existing TowerScout application to utilize a more enterprise architecture on Databricks. It's ideal for users of TowerScout who are looking for a more scalable system
This migration introduces the following features:
- Stored detections in ADLS/S3
- Retraining pipeline connected to validated images
- Model scaling and storage in Unity Catalog
We do not provide the model in this release; however, YOLOv5 and EfficientNet model weights available upon request, you can train your own model using the training pipeline.
We developed this using DBR 15.4 ML. Although this application was designed on Databricks, you can adapt the code to a standard Spark setup
The original TowerScout is monolithic and that may serve some use cases better, you can find that repo here.
The frontend files are located in webapp
The backend/ML files are located in tsdb
Notebooks are the orchestration tools we used for the PySpark backend. The application runs using file arrival to trigger a structured streaming query. We have a timer listener that closes the stream after a period of time.
We run this using YOLOv5, but if you opt to run this on Databricks, using YOLOv10 along with a ML Endpoint will yield better performance.
The source code in this repository is licensed under the Apache License, Version 2.0. See the LICENSE file for details.
This project integrates with proprietary cloud services including Microsoft Azure Maps and Azure Databricks. Use of these services requires a valid subscription and acceptance of their respective terms of service.
No Azure Maps credentials, Databricks credentials, or proprietary Microsoft or Databricks software are included in this repository.