Object-Centric Representation Learning
Overview
Object-centric representation learning asks how visual systems can represent a scene as separate entities, rather than as one undifferentiated image, so they can track objects, reason about relations, and predict what happens next. Common approaches use slot attention and reconstruction objectives, self-supervised feature grouping, or temporal and 3D models that separate persistent identity from changing appearance and motion. Researchers evaluate these representations on object discovery and video segmentation benchmarks such as CLEVR and MOVi, as well as compositional reasoning and robot manipulation tasks such as LIBERO. A central tension is that object structure can improve generalization and control under visual or compositional shifts, but slots may fragment or bind objects poorly, and dense pretrained features can be stronger when tasks are easier or training data is abundant.
Latest papers
94 papers11 papers in the last four weeks, up 22% on the four weeks before. 0.1% of all new papers.