An idea on how this could work: repeatedly re-split the dataset (to cover all of it), and re-train a detector on the splits, then at the end of each training cy

An idea on how this could work: repeatedly re-split the dataset (to cover all of it), and re-train a detector on the splits, then at the end of each training cycle surface validation frames with the highest computed loss (or some other metric more directly derived from bounding boxes, such as the number of high confidence "false" positives which could be instances of under-labeling) at the end of training. That's what I do on noisy, non-academic datasets, anyway.