Are Sampling Heuristics Necessary in Object Detectors?

To address the imbalance between foreground and background, various heuristic methods, such as OHEM, Focal Loss, GHM, have been proposed for biased sampling or weighting when training deep object detectors. We challenge this paradigm by discarding the sampling heuristics and focusing on other settings for training. Our empirical study reveals that the weight of classification loss and the initialization strategy have a big impact on the training stability and the final accuracy. Thus, we propose the\emph {Sampling-Free} mechanism, including three key ingredients: optimal bias initialization, guided loss weights, and class-adaptive threshold, for training deep detectors without sampling heuristics. Compared with the sampling heuristics, our Sampling-Free mechanism is fully data diagnostic and thus avoids the laborious tuning of sampling hyper-parameters. Our extensive experimental results demonstrate that the Sampling-Free mechanism can be used for one-stage, two-stage, and anchor-free object detectors, where it always achieves higher accuracy on the challenging COCO benchmark. The mechanism is also useful for the instance segmentation task. Code is at this https URL.