Wenhui Chu

SAM struggles in robotic vision not because of the entire model, but because of representation shifts in shallow layers, and RepSAM exploits this to achieve near-full fine-tuning performance with just 0.6% of the parameters.