Most models that mix vision, audio, and language are good at telling you what's in a scene, but not where it is. They'll recognize a dog and a passing car yet lose track of the fact that the dog is on your left and the car behind you. SceneBind adds that spatial sense. It represents a scene with one overall semantic embedding plus a set of object slots that each carry meaning, position, and even a measure of uncertainty.

To make that useful, it comes with a matching scheme combining whole-scene similarity with object-by-object alignment, which supports retrieving scenes or grounding specific objects across modalities. They also built a real-world binaural audio-visual dataset (binaural meaning two-ear recording that carries direction) to train and test on. The spatial part is deliberately lightweight, riding on top of existing semantic encoders with only a few extra tokens.

All of this is the abstract's telling of their own results, so the paper holds the datasets, baselines, and numbers.