We just shipped the biggest upgrade to our pin detection model since we launched it: v4, trained on a dataset of over 154,000 images. That's roughly 9x the size of what we used for v3. The dataset is a mix of two things: 151,000 patches sliced from 11,500+ real scraped pin board photos from across the internet, and 3,200+ synthetic boards we generated ourselves using our library of 155,000 individual pin images. The synthetic data was a bit of an experiment — creating artificial board images by compositing individual pins — and it paid off. Blending real and synthetic gave the model much better generalization across different board layouts, lighting conditions, and pin densities.
For training we had access to an NVIDIA RTX PRO 6000 Blackwell Server Edition with 96GB of VRAM. This is a very new GPU — it just launched — and getting to train on it was genuinely exciting. The extra VRAM headroom let us run larger batch sizes and go the full 100 epochs without memory constraints getting in the way. We trained YOLO11m (the latest generation) and ended up with an mAP50 of 0.953 — that's 95.3% detection accuracy — with precision of 0.904 and recall of 0.899. Those numbers represent a significant jump over v3 across all three metrics.
The v4 model is already deployed and live on the site. If you've run an upload through the AI detection pipeline in the last day or two, you've already been using it. In practice the improvement is most noticeable on dense boards where pins are closely packed together, and on boards photographed in dim or inconsistent lighting. The model has seen enough variety in the training data that it handles edge cases much more gracefully than before.
We're a small hobby project and we don't have the resources of a big company, but we believe in using the best tools available when the opportunity comes up. Good training data and good hardware are the main levers you can pull in this kind of work, and this round we got to pull both of them hard. The detection side of the pipeline is in a really good place now — next focus will be improving the matching accuracy on the CLIP side for tricky pins that look similar to each other.