arxivcs.CV2026-07-09
HSA: Hierarchical Slot Attention for Multi-granularity Scene-Decomposition
Neelu Madan, Rongzhen Zhao, Andreas Mogelmose, Juho Kannala, Joni Pajarinen, Graham W. Taylor, et al.
Slot attention is a powerful framework for object-centric learning, decomposing visual scenes into latent slots through iterative competitive attention. However, existing methods share two critical limitations: they decompose scenes into a flat set of slots at a single granularit…