What's this about?
Rethinking efficient generative spatio-temporal video grounding — finding *where* and *when* something described in text occurs in a video. 15 upvotes.
What is video grounding?
Given "the moment the dog jumps onto the sofa", return the time span and the region of frame it happens in. It is harder than image detection in both dimensions: you need the temporal boundary of an event, which is genuinely ambiguous, and the spatial region within it. Doing this *generatively* — producing the answer rather than scoring candidate regions — is the efficiency argument, because scoring every candidate span in a long video is prohibitive.
Why does this matter to you?
- Grounding is the capability that turns video from something a model describes into something it can be questioned about precisely.
- It is the retrieval layer for any agent that needs to work with video evidence.
- Efficiency is the whole contribution — the accuracy problem was already partly solved at costs nobody wants to pay.