Figure 10

From one slot read to a full mask

1 · Photo + phrase

“the man with his arms out”

2 · One slot read

box (cx, cy, w, h) and the mask centre, read at once
~146 ms read

3 · Prompt SAM 2

the box grown 20% + the centre point
+~20 ms per image

4 · Mask

SAM 2's mask (filled) vs our 24-ray polygon (dashed)
polygon alone 0.56 → SAM 2 0.79 mIoU

mIoU on 1,000 RefCOCO val objects (COCO val2017: 0.54 → 0.76). SAM 2 from the true box reaches 0.86, so most of the remaining gap is the box. Growing the box matters: SAM 2 won't draw outside the box it's given, and boxes a little too tight crop the object. The SAM 2 mask in stage 4 is drawn as the true outline for illustration.