Revisiting Foreground-Background Imbalance in Object Detectors

We report comparable COCO AP results for object detectors with and without sampling/reweighting schemes. Such the schemes, e.g. undersampling, Focal Loss and GHM, have always been considered as an especially essential component for training detectors, which is supposed to alleviate the extreme imbalance between foregrounds and backgrounds. Nevertheless, our report reveals that for addressing the imbalance to achieve higher accuracy, these schemes are not necessary. Specifically, by three simple training/inference strategies - decoupling objectness from classification, biased initialization, threshold movement, we successfully abandon sampling/reweighting schemes in the representatives of one-stage (RetinaNet), two-stage (Faster R-CNN), and anchor-free (FCOS) detectors, with the not worse performance than the vanilla models. As the sampling/reweighting schemes usually introduce laborious hyper-parameters tuning, we expect our discovery could simplify the training procedure of object detectors. Code is available at https://github.com/ChenJoya/resobjness.