I noted that the extreme‑point method increased mean Intersection‑over‑Union (mIoU) from 68.2 % to 77.5 % and pixel accuracy from 82.1 % to 90.3 % on occlusion‑heavy benchmarks.
The approach detects extreme points on each object, extracts embeddings, computes pairwise cosine similarity, and propagates labels via graph diffusion to recover occluded regions.