Active learning that targets a geospatial model's blind spots
Problem
In edge-case regions, models can be confused by natural features and mistag objects like yard debris or trees, leading to inaccurate predictions. Fixing those edge cases is complex, requiring hours of coordination across many stakeholders.
Solution
Cape ran an iterative active-learning cycle inside its training pipeline, finding low-confidence areas and prioritizing them for additional signal via Labelbox. The cycle increased model accuracy in a targeted way; specifying where the model was uncertain allowed rapid fixes and more performant models.
Result
The aggregate time saved from Labelbox's active-learning workflows represented an estimated 30%+ increase in total time savings, plus months of custom development work in engineering hours.
Cape Analytics extracts property attributes from geospatial imagery for insurers. Labelbox's active-learning loop targeted the model's low-confidence cases, saving 30%+ time and months of engineering.
The challenge
Cape Analytics gives insurers and property stakeholders valuable property attributes during underwriting, using computer vision to extract information from geospatial imagery. It combines inspection-level detail with the speed of a living database covering the entire US property base, delivered in seconds. Building those models is hard where the taxonomy is subtle. Identifying yard debris matters — it raises insurance risk — but debris is hard to distinguish from unarranged yard furniture or construction materials, and in some regions models confused natural features for debris. The team needed to find and fix exactly those failure cases.
The approach
Cape used Labelbox to run an iterative active-learning cycle inside its training pipeline. Low-confidence predictions were surfaced and prioritized for data scientists and contributors, who analyzed and corrected them — increasing model accuracy in a targeted way. Specifying where the model was uncertain let the team fix the right things fast. Routing those low-confidence cases to the right people took well-designed queue management, which Labelbox provided.
There are many labeling tools out there but the Labelbox backend is the real differentiator. With dynamic queueing, our labelers are never out of work, which was a major upgrade compared to our old internal tools — both from a productivity and speed-to-production point of view.
— Cape Analytics, Head of Engineering
The outcome
The combination of queue management and active-learning cycles saved an estimated 30%+ in total time, plus months of custom development work in engineering hours. Cape's targeted approach to defining risk is a key reason it's growing fast, applying rigorous deep learning to geospatial imagery for insurance and beyond.
Where this goes
A model that flags its own uncertainty, fed signal exactly where it's weakest, improves faster than one trained on more of everything. That's the efficient path to production accuracy.