Remote Visual Inspection
How to bring together 9 data scientists working on 5 Computer Vision models & 1 million images to train a visual inspection model and deploy it into production for the first time.
ENERGY
LEAD DATA SCIENTISTS
SCALE PRODUCTION
Context
The data science team at a large energy operator in APAC decided to use computer vision to improve the process of remote visual inspection of power grids.
Images are collected by drone operator teams and foot patrols in the fields. They are manually reviewed by trained engineers to identify defects. The footage from a single tower generates GBs of data, takes days to analyze, and is prone to error.
With minimal resources - an open-source data versioning tool, GPU, data storage, serving infrastructure, and notebooks - the DS team completed an initial POC. They then needed to evaluate the benefits of using CV for review teams in a 3-month pilot with new field data sets.
From 1 month to 1 week to iterate on a model.
+33% of time spent by data scientists on high-value tasks.
From 100% to 50% of images manually reviewed. Only 30% of false positives.
Datalake: a centralized datalake to index, search and store all the data and track sources and metadata.
Monitoring System: a model & data monitoring system to assess quality & performance.
Model Registry: a single point of truth to version and centralize all models.
Hybrid deployment: Store data and deploy models on their own infrastructure. Leverage the Picsellia GDPR infrastructure for model training.
