Paper Review: "OTA: Optimal Transport Assignment for Object Detection"

As we already have the cost matrix, supplying vector s and demanding vector d, the optimal transportation plan π* can be obtained by solving this OT problem via the off-the-shelf Sinkhorn-Knopp Iteration. Noted OTA only increases the total training time by less than 20% and is totally cost-free in testing phase. Previous methods only select positive anchors from the center region of objects with limited areas, called Center Prior. For general detection datasets like COCO, the authors find the Center Prior still benefit the training of OTA. Hence, they impose a Center Prior to the cost matrix. For each gt, they select r 2 closest anchors from each FPN level according to the center distance between anchors and gts. As for anchors not in the r 2 closest list, their corresponding entries in the cost matrix c will be subject to an additional constant cost to reduce the possibility they are assigned as positive samples during the training stage.