Although the current popular object detection algorithms have quite different design paradigms, they indeed have some common points, that is, almost all object detection algorithms contain a candidate region proposal module and a boundary box regression (BBR) module. The candidate region proposal module is used to find some candidate regions which may (or may not) contain objects, while the BBR module is used to fine-tune the bounding boxes of the candidate regions to obtain more accurate locations. For example, the R-CNN algorithm uses selective search [
15] to extract about 2000 candidate regions for each image; faster R-CNN uses region proposal network (RPN) to find candidate regions; YOLO and SSD use grids and anchors to generate a set of candidate regions, respectively. Candidate region proposal module is introduced to avoid the huge search space of sliding windows, which may cause huge computation cost. It should be noted that no matter which kind of candidate region proposal technique is used, all of these object detection algorithms include a BBR module in the final stage of the detection process to refine the position of object. BBR was first proposed and used by Felzenszwalb et al. [
16], and then used in almost all object detection algorithms [
17]. BBR module can work alone, or be integrated with other modules.