Co-designed and co-implemented a fully classical computer vision pipeline to detect and classify UNO cards in real-time game images: identifying each card's color and value and mapping it to the correct player region, with no neural networks involved.

• Region Layout: Defined the five player/center polygons and tuned their vertex coordinates iteratively to maximise region accuracy across the test set.
• Active Player Detection: Implemented both token detectors: the dark rectangular blob on white backgrounds and the yellow circular blob on dark backgrounds, including the background-type classifier that selects between them.
• Contour Filtering & Cropping: Detected blobs from the HSV masks, approximated each to a convex polygon, filtered out candidates that were too small or too large, merged overlapping detections, and extracted the bounding box crop of each remaining candidate for classification.
• Post-Processing & Region Assignment: Handled all ambiguity resolution after template matching: deduplicating multiple detections of the same card, resolving the draw-2 / 2 confusion, applying lower confidence thresholds for black (wild/draw-4) cards, and assigning each detected card to its correct player region via pointPolygonTest depth.