SPOT: Self-Training with Patch-Order Permutation for Object-Centric Learning with Autoregressive Transformers

4 IACM-Forth 5 Archimedes/Athena RC Figure 1. SPOT: Our novel framework enhances unsupervised object-centric learning in slot-based autoencoders using self-training and sequence permutations in the transformer decoder. It improves object-specific slot generation, excelling in complex real-world images.