← Lecture

LANGUAGE-GROUNDED VISION

Change the words.
Inspect the region.

One controlled scene, five spatial interfaces. Compare authored targets with recorded model predictions.

Authored park with a dog, a child kicking a red ball, a blue ball and a tree

SPATIAL OUTPUT

Provenance and model settings

Recorded predictions are replayed offline. No live model is loaded by this page. Grounding scores and SAM quality estimates are model scores, not measured accuracy.

Download all recorded outputs · Download the recording kit · View the notebook
Geometry sandbox: move a box, watch IoUIoU = 0.600