BBRefinement: A Universal Scheme to Improve the Precision of Box Object Detectors

We present a conceptually simple yet powerful and general scheme for refining the predictions of bounding boxes produced by an arbitrary object detector. Our approach was trained separately on single objects extracted from ground truth labels. For inference, it can be coupled with an arbitrary object detector to improve its precision. The method, called BBRefinement, uses a mixture of data consisting of the image crop of an object and the object's class and center. Because BBRefinement works in a restricted domain, it does not have to be concerned with multiscale detection, recognition of the object's class, computing confidence, or multiple detections. Thus, the training is much more effective. It results in the ability to improve the performance of SOTA architectures by up to two mAP points on the COCO dataset in the benchmark. The refinement process is fast; it adds 50-80 ms overhead to a standard detector using RTX2080; therefore, it can run in real time on standard hardware. Finally, we show that BBRefinement can also be applied to COCO's ground truth labels to create new, more precise labels. The link to the source code is provided in the contribution.