We've been quietly working on something we're genuinely excited about: a custom YOLO-based object detection model trained specifically for Disney pins. This isn't a generic image recognition tool — it's a model we built and tuned ourselves, running entirely on our local mini server (an AMD 7840HS with a Radeon 780M integrated GPU). The fact that this runs on modest consumer hardware without needing a cloud GPU is something we're pretty proud of.
The model can detect individual pins from photos of boards, lanyards, and pin cases. That's the hard part — figuring out that there are 40 separate pins in a single photo of someone's board, and drawing a bounding box around each one. Once we have those individual pin crops, we run them through a CLIP visual search system that matches each detected pin against our database of 155,000+ reference pin images. The result: you upload a single photo of your pin board, and we come back with identification attempts for every pin on it.
Accuracy is still improving — pin detection is genuinely tricky, especially on crowded boards with overlapping pins, weird lighting, or pins photographed at an angle. But even at this stage the results are useful, and we're actively collecting feedback to keep making it better. If you try it and it misses something obvious or makes a weird call, let us know — that feedback directly shapes how we improve the model.
The whole pipeline — detection, cropping, embedding, search — runs locally on our little server. No cloud GPU required for inference. It's slow enough that we queue jobs rather than processing them live, but the quality of the matches makes it worth the wait. This is one of those features that started as a "wouldn't it be cool if..." and turned into something we actually use ourselves every time we add pins to a collection.