we investigate the scale alignment between pre-training and target datasets, and propose a new refined Scale Match method (termed SM+) for tiny person detection. SM+ improves the scale match from image level to instance level, and effectively promotes the similarity between pre-training dataset and target dataset.
While SM only considers the whole image, our SM+ focuses on every instance. The instance-level approach achieves scale match in a finer level. Our SM+ mainly consists of four steps: (1) Separation, (2) Instance processing, (3) Background processing, and (4) Combination.
Background based on inpainting (top) vs. Background based on new sampling (bottom). The inpainting method might not repair some artifacts, but changing the background does not cause this problem.