Falcon Perception: 0.6B Early-Fusion Transformer for Grounding and Segmentation

AI Impact Summary

Falcon Perception introduces a novel 0.6B-parameter early-fusion Transformer model designed for open-vocabulary grounding and segmentation. The architecture’s key innovation is a hybrid attention mask and Chain-of-Perception output interface, enabling variable-length dense instance predictions. This model demonstrates strong performance on the SA-Co benchmark, particularly in spatial understanding and object detection, and introduces PBench, a diagnostic benchmark to isolate capability gaps, offering a structured approach to model improvement.