The pycoco_wrapper only evaluates on images where there is a prediction, and leave out image_ids where there is not prediction.
This is a serious bug if I understand it correctly. The coco metrics should be evaluated on all the images.
Correct way to do this should be below, without filtering image_ids.

    coco_dt = coco_gt.loadRes(predictions=coco_predictions)
    coco_eval = COCOeval(coco_gt, coco_dt, iouType="bbox")