My safety-gear detector reported 0.91 mAP and a −93% false-positive cut. Then I audited my own evaluation and found three validity problems — two data leaks and a seed confound — that make both numbers softer than they sound. Here's the one I'd actually trust.
RTMDet trains great on one consumer GPU — once you survive the prebuilt-wheel lock, a pkg_resources crash, and a pip ResolutionImpossible (and debunk the SyncBN scare). The exact pins and fixes that worked.
A review of RTMDet (arXiv 2022): a real-time detector that reaches 52.8% COCO AP at 300+ FPS through systematic design choices rather than a new module, and extends cleanly to instance segmentation and rotated detection.
A safety-gear detector scored 0.91 mAP yet fired 'helmet_off' on helmeted workers. Here is the diagnosis, and how empty-GT negatives cut external false positives by 66% with no measured loss on the AIHub in-clip validation mAP.
Solution to Grid Adjacent Danger Zone Detection
Optimizing PlantUML diagram layout from a crossing-prone grid to a clean vertical stack.
A classical image processing baseline reached IoU 0.9795 on 5 MRI slices, but the result came with important validation caveats.
Common pytest fixture patterns for clean, maintainable, and reusable test code.
How to document API test scenarios effectively using diagrams and structured formats.
Practical tips for using the Markdown PDF extension in VS Code, including common pitfalls with Mermaid diagrams.