Improving Panoptic Segmentation at All Scales

Panoptic segmentation on high-resolution natural images is challenged with recognizing objects at a wide range of scales. Standard approaches (left) can struggle when dealing with very small (zoomed detail) or very large objects (bus on the left). By introducing a novel instance scale-uniform sampling strategy and a crop-aware bounding box loss, we are able to improve panoptic segmentation results at all scales (right).