Hey, I'm currently developing an object detector and tracker for an autonomous driving application as part of my master thesis. I built my own dataset with 4,5k images containing the objects I want to detect and I'm already quite satisfied with my results compared to the dataset size (~90 % AP for the classes i care most about). Nevertheless I still have a good amount of False Positives when evaluating my Yolo v4 based detector on unseen (video)data. My object tracker can already mitigate some of these FPs (e.g. when they only occurred in one frame) but I was wondering if there are other worthy strategies to further mitigate this problem. Surely I could simply increase the dataset size or increase the network input size, but I'm looking for ideas/strategies beside that. I'd be grateful for some tips, ideas or papers that are worth reading :)