Figure 9

Boxes and masks are just more slots

A box is a centre and a size: 4 slots. A mask is a centre and 24 ray lengths, how far the object reaches in each direction: 26 slots. Same template idea, same single read. Sizes and ray lengths are on a log scale, since objects span three orders of magnitude.

The mask centre is the point deepest inside the object (the maximum of its distance transform), so it is always inside. Each ray records the farthest object pixel along its direction. Rays are 15° apart, so thin parts between rays (an arm, the gap between the legs) are cut off. The polygon alone reaches ~0.56 mIoU on RefCOCO; handing the box and centre to SAM 2 gets 0.79.