YOLO11n Hits mAP, Falters in Practice

The model is probably not “bad” so much as optimized for the wrong target: mAP50 on a validation set is a weak proxy for field performance.

• –YOLO11n is a nano model built for edge constraints, so it can struggle with small objects, occlusion, clutter, and fine-grained classes.
• –If the dataset is narrow, noisy, or missing hard negatives, the model can score well on paper and still fail on live camera input.
• –Practical evaluation should use the real deployment setup: camera angle, distance, lighting, motion blur, frame rate, and the precision/recall threshold you actually care about.
• –For a CPU-only Raspberry Pi, the biggest wins usually come from better labels, more representative data, and sensible image-size tradeoffs before changing architectures again.
• –If the task still needs stronger performance, the bottleneck may be hardware or model capacity, not training technique alone.