When Slots Compete: Slot Merging in Object-Centric Learning

A team of researchers has introduced a novel technique called 'slot merging' to address a fundamental flaw in object-centric learning models. These models, like the established DINOSAUR pipeline, represent an image as a set of latent 'slots' that are decoded into features. A persistent issue is that with a fixed number of slots, multiple slots often end up competing for and representing overlapping parts of the same object, rather than cleanly separating distinct entities. This competition degrades the model's ability to factor scenes into their constituent parts.

The proposed slot merging acts as a drop-in, lightweight operation during training. It quantifies the overlap between slots using a Soft-IoU (Intersection over Union) score calculated from their attention maps. When a pair of slots exceeds a dynamically inferred overlap threshold, they are merged using a barycentric update that preserves gradient flow for stable training. Crucially, the entire process follows a fixed policy and requires no additional learnable modules, keeping it computationally efficient.

Integrated into the DINOSAUR framework, which uses feature reconstruction, slot merging demonstrably improves object factorization and the quality of the segmentation masks the model produces. The authors report that their method surpasses other adaptive techniques on standard object discovery and segmentation benchmarks. This represents a significant step toward more interpretable and reliable computer vision systems that can autonomously decompose complex scenes.