Focus bank: an innovative mechanism for improving the performance of focus-and-detect algorithms in tracking multiple soccer players

The goal of multi-object tracking(MOT) is to detect and track interested objects in videos. It is more challenging to track multiple players in soccer videos, and most existing detection and tracking approaches are not applicable owing to the factors such as similar appearances. Focus-anddetect approaches are introduced to soccer scene in this paper. These approaches, originally proposed for small objects, comprise two stages: the focus stage predicts the regions that potentially contain objects using low-resolution feature maps, while the detect stage performs subsequent detection task based on the results of focus stage. Furthermore, an innovative mechanism focus bank is proposed to complement the focus stage results in tracking by collecting and analysing the information from previous frames. QueryDet is chosen as the representative of the focus-and-detect approaches and conducts extensive experiments on the SoccerNet dataset. In order to demonstrate the flexibility and generalizability of the focus bank, two focus bank schemes are designed and applied to three different trackers respectively. The experiments results show that these approaches achieve consistent improvements in MOTA ranging from 0.1 to 0.4 for all trackers, only with negligible computational cost.