Deep learning pipeline trained on Detectron2 to detect and measure kitchen elements from architectural elevation drawings — combining elevation detection, OCR-based dimension extraction, pixel-to-real-world scaling, and NMS-based conflict resolution.
Architectural elevation drawings are highly domain-specific, contain dense overlapping elements, and embed dimensional information as handwritten or printed text annotations rather than metadata — making standard object detection approaches insufficient without a custom multi-stage pipeline combining detection, OCR, and geometric reasoning.
Built a sequential pipeline using Detectron2 trained on custom elevation annotations. First gated inputs through an elevation classifier, then used OCR to parse dimension text and establish a pixel-to-real-world scale ratio, then ran element detection within validated elevation regions, and finally applied NMS and conflict resolution to produce clean, dimensionally accurate kitchen element detections.
Engineered a specialized computer vision pipeline for automated detection and dimensional analysis of kitchen elements from architectural elevation view drawings — a domain where standard object detection models fail due to the highly structured, technical, and annotation-sparse nature of architectural CAD imagery.
The pipeline begins with an elevation detection stage: given an input drawing, the system first classifies whether elevation views are present and identifies how many exist within the document. This gating step ensures downstream processing is only triggered on valid elevation content, avoiding false detections on plan views or other drawing types.
OCR is then applied to extract dimension annotations written near the elevation views — textual measurements such as widths and heights noted in standard architectural notation. These values are parsed and used to establish a real-world-to-pixel scale ratio, enabling the system to derive accurate physical dimensions from pixel measurements within the drawing.
With the scale calibrated, the Detectron2-based element detection model — trained on custom-annotated elevation view datasets — identifies individual kitchen elements such as cabinets, appliances, and fixtures within each elevation. Each detected element is assigned its real-world width derived from the pixel scale. Non-Maximum Suppression (NMS) and additional post-processing methods are applied to resolve overlapping detections and refine bounding box accuracy in dense elevation layouts.
Tangible Impact
Delivered an end-to-end automated kitchen element detection system capable of processing architectural elevation drawings and returning spatially accurate, dimensionally annotated element detections — significantly reducing the manual effort required in kitchen design and estimation workflows.
© 2024 NIKHIL
BACK TO TOP ↑