Object-detector misses should be triaged from low-threshold proposals before changing data or…

Object-detector misses should be triaged from low-threshold proposals before changing data or architecture. A well-localized low-score box suggests calibration or classification, a misaligned box suggests localization or label-policy problems, and no nearby proposal suggests representation or scale limitations. These categories need different experiments; lowering one operational threshold can recover some misses while sharply increasing false positives. This method is most useful when proposal geometry, confidence, and the operating gate are all available, and conclusions from small validation sets should remain hypotheses until reproduced on an untouched holdout.