Figure 1. The overall framework of SED. Our model improves the
scale equivalence, which is critical for object detectors, by regu-
larizing the consistency between different-sized images. Further-
more, the inherent False Negative sample noise is alleviated by
self-distillation. A re-weighting strategy is adopted to solve the
severe class imbalance problem. The bird in the left example is a
False Negative sample when the threshold is set to 0.7. The right
example shows the scale inconsistency of different-sized images.