Hi @picekl, if you look at our first submission, we had got a good overall_precision score but a lower overall_recall score according to the evaluation metric. Therefore, we improved our model to get a better overall_recall score, which we achieved with our subsequent submissions.
But, by then, when we submitted, the ‘overall_recall’ column was removed. After getting confirmation from @dimitri.fichou via mail that overall_precision score is going to be the sole evaluation metric for the competition, we re-trained our model to improve on the precision scores. This is reflected in our last two submissions.
We believe that even if the evaluation metric is modified to consider either the f1 score or mAP (over IoU > 0.5), two of our submissions would excel in that, as they were trained particularly to increase the same.
@dimitri.fichou, it would be great if you could clarify what exactly would be the final evaluation metric. We’ll make another submission and tag that as the ‘Primary Run’.