Object-centric learning (OCL) uses slots to extract object representations, enhancing flexibility and interpretability. Slot attention, a common OCL method, refines slot representations with attention mechanisms but requires predefined slot numbers, ignoring object variability. To address this, a novel complexity-aware object auto-encoder framework introduces adaptive slot attention (AdaSlot), dynamically determining the optimal slot count based on data content through a discrete slot sampling module. A masked slot decoder suppresses unselected slots during decoding. Extensive testing shows this framework matches or exceeds fixed-slot models, adapting slot numbers based on instance complexity and promising further research opportunities.
0 commit comments