We only have one class, human, in our dataset. The quality of the bounding boxes will be evaluated by computing [Average Precision (AP)](https://en.wikipedia.org/wiki/Evaluation_measures_(information_retrieval)#Average_precision). However, it seems like the Kaggle Community Competition doesn't support calculating AP for object detection, we have to use Absolute Error (AE) instead. On this other hand, I think using F1 score is easier and see the influence of different confidence threshold. Therefore, you only need to submit your F1 score at your selected threshold. The Leaderboard will show the error (1 - your_AP).