How Computer Vision & YOLOv8 Identify 15,000+ LEGO Minifigures in Milliseconds

Identifying a specific LEGO minifigure with the human eye can be frustrating. A classic Star Wars stormtrooper might look identical to a novice, but minor variations in helmet mold year, torso print line thickness, or leg dual-molding can mean the difference between a $4 common figure and a $350 rare variant.

Here is how the HelloBrick computer vision engine solves this challenge in real time.

---

The Computer Vision Pipeline

Our inference engine runs a multi-stage convolutional and transformer pipeline:

1. Video Stream (60fps)
       │
       ▼
2. Spatial Segmentation (SAM / YOLOv8-Seg) 
       │  Detects bounding box + contour isolation
       ▼
3. Feature Extractor (High-Contrast Cropping)
       │  Torso pattern, head print, accessory tokens
       ▼
4. Vector Embedding Distance Match (Cosine Sim > 0.94)
       │  Queries 15,000+ item embeddings
       ▼
5. AR HUD Projection (Price + Condition Overlay)

---

Handling Glare and Specular Reflections

Plastic ABS bricks are notoriously reflective. Harsh overhead lighting creates specular highlights that obscure surface printing.

To counteract this, HelloBrick employs:

• Adaptive Histogram Equalization (CLAHE): Balances dynamic range in high-glare environments.
• Invariant Feature Matching: Keypoint descriptors focus on geometric lines (e.g. belt buckle prints, helmet visor angles) rather than raw color saturation.

---

The Result: Sub-100ms Recognition

By optimizing our YOLO models and running direct vector lookup against Supabase pgvector and edge endpoints, HelloBrick achieves recognition latency under 85ms on modern iPhones.

Want to see it in action? Grab any minifigure on your desk, open HelloBrick, and hover the camera!