Hi @zindi! Can you please share evaluation metric - because right now I have strong assumption that metric is different from standard mAP metric and have some separate logic for NEG class - I have submitted 3 submits with different lower threshold levels for confidences (0.1 0.2 0.05) and scores differs a lot and if we follow standard COCO eval from pycocotools the difference is such small thresholds should not be so high.