Perceptio

Large vision-language models describe images well but are vague about where things are. Perceptio makes the model emit segmentation and depth tokens first, then answer. Short note on the idea, the training tricks and the results.