Cycle Consistency in Video Object-Centric Learning

Researchers propose implicit cycle consistency for self-supervised video object-centric learning, aiming to improve object discovery and temporal association.

The paper argues that explicit cycle consistency from multi-object tracking does not fit stochastic object-centric slots and can cause feature collapse.

Instead, it enforces consistency on the reconstruction manifold, which better matches the video OCL setting.

The method reports stronger results on video benchmarks for object discovery and tracking across time.