Slot Attention

Slot attention uses a small set of learned slots that compete to explain visual features, often trained with reconstruction losses to discover objects without object labels. Variants add temporal memory to preserve object identity in video, adapt the number of slots, or impose hierarchical and geometric structure; related methods also apply slots to audio-visual scenes, robotics, and medical images. Evaluation commonly measures object segmentation and tracking on synthetic and real-world benchmarks such as MOVi and YouTube-VIS. A central challenge is making slot assignments reliably capture meaningful objects and persist through occlusion, while deciding how much supervision or structural bias is needed.