Detecting utility defects from drone imagery with CV signal
Problem
The Sharper Shape team previously used heavily manual workflows and open-source data processing tools that lacked the configuration and customer support its needs and SLAs required.
Solution
Labelbox Annotate let Sharper Shape connect raw data via a simple API. Collaboration features enabled rapid onboarding, training, and throughput for internal and external contributors to produce signal together in one centralized environment.
Result
Sharper Shape now concentrates on model building and deployment without extra engineering effort, and sped up model training by over 10X. It also cut average costs to create training data by as much as 50% while maintaining the highest-quality signal.
Sharper Shape's drones run computer vision to spot utility hazards. Labelbox produced the image signal and a model-assisted loop, cutting labeling cost 50% and speeding training 10x.
The challenge
Sharper Shape builds technology for safe, efficient utility transmission and distribution, using drones to inspect power infrastructure. Its computer vision models, running in advanced aerial sensor systems, automatically collect and analyze inspection data — flagging dangerous setups like vegetation growing too close to wires or broken insulators so utilities can address hazards. Training multiple computer vision models takes a vast amount of accurately graded imagery. Before Labelbox, the team relied on heavily manual workflows and open-source tools that lacked the configuration and support it needed.
The approach
Sharper Shape used Labelbox to produce the signal and streamline its pipeline, connecting raw data via a simple API and working across data types like tiled imagery in one place. Collaboration features let internal and external contributors onboard, train, and produce signal together in one centralized environment. The team then accelerated with model-assisted feedback — importing its model into Labelbox and focusing effort on edge cases.
With the streamlined design of Labelbox, we are able to cut costs on data processing by as much as 50% while maintaining the highest quality in our training data, and get to training our models faster. With model-assisted expert feedback, we expect another huge reduction in time and costs to process. After a preliminary model is trained, we can run a loop to generate signals from our model’s inference, and feed those back into Labelbox, effectively cutting the processing load of our labelers to that of reviewing for false positives. That allows us to increase our capabilities and model accuracies exponentially with respect to time for the amount of components and defects we can detect and classify.
— Edward Kim, Data Analyst / AI at Sharper Shape
Before using Labelbox, we struggled with managing our training data creation infrastructure and manual experiment tracking. We’re now able to concentrate on model building and deployment, without sparing engineering effort and are able to speed up model training by over 10X.
— Jaro Uljanovs, Data Science / AI at Sharper Shape
The outcome
Sharper Shape cut average cost to create the training data by as much as 50% while maintaining the highest-quality signal, and sped up model training by over 10X — freeing the team to focus on model building and deployment. With model-assisted human feedback loop, contributors increasingly review for false positives instead of grading from scratch, expanding the components and defects the models can detect and classify.
Where this goes
Grid inspection is embodied AI in the air. A model-assisted loop, where the model produces signal and experts review it, is what scales detection across a continent of infrastructure.