01.
Window photographs
Submit the window photographs.
Loading the pageLoading the pageLoading the pageValidating computer vision for window recognition
ACW is developing CleanQuote’s window-recognition workflow and a validation interface for comparing model output with checked reference results.
CleanQuote turns window photographs into section counts and area estimates for a quoting workflow. The two outputs solve different tasks: the system needs to identify the sections in an image and estimate the area they represent. Each output therefore has its own validation measure.
CleanQuote turns window photographs into section counts and area estimates for a quoting workflow. ACW’s computer vision work treats these as separate outputs, with detection and filtering refined to distinguish the relevant window sections from extra detections. The result: computer vision outputs connected to a practical quoting workflow.
Submit the window photographs.
Identify the relevant window sections.
Return section counts and area estimates for review.
Discovery established the validation approach and architecture before model and validation-interface development. Section counts and area estimates are evaluated separately for model cleanquote-engine-1.7.4. The validation cards below show the current results and the request-level timing.
Evaluate section-count and area-estimation outputs separately.
Compare the outputs with the evaluation records.
Review accuracy and request-level processing time.
Separate, reviewable measurements for section-count accuracy, area-estimation accuracy and request processing.
The current model achieved 99.4% section-count accuracy.
Model cleanquote-engine-1.7.4; 334 evaluation records.
The current model achieved 97.78% area-estimation accuracy.
Model cleanquote-engine-1.7.4; 334 evaluation records.
Requests containing up to 10 window photos were processed in 19.4 seconds on average.
The wider dataset contains more than 3,000 photos.
A model output becomes useful when the team can inspect it in the quoting process. ACW connects the vision engine with a web workflow for reviewing section counts and area estimates, keeping the submitted material and the returned measurements in the same product context. The result: a reviewable product workflow around the vision engine.
