Releasing my first paper, Moondream Segmentation: From Words to Masks
We train Moondream to produce boundary-accurate segmentation masks by first autoregressively generating coarse masks and then iteratively improving them with dedicated refiner layers. We achieve 80.2 cIoU on RefCOCO val and 87.6 cIoU on RefCOCO-M.
Big thanks to @vikhyatk and the rest of the @moondreamai team