Falcon Perception-HD GRPO update detects 100-500 objects, dropping NMS entirely

HildeKuehne · x · 2026-09-26

Falcon Perception-HD will be presented at NeurIPS 2026. The GRPO post-trained weights markedly improve referring expression comprehension and high-density perception (detecting 100-500 objects), while fixing maximum-likelihood failure modes like repetition, low recall and object merging — eliminating post-processing heuristics such as NMS. Paper, weights and a playground are public.

Original post →

More from Multimodal

Multimodal channel →