Confidence Matters: Revisiting Intrinsic Self-Correction Capabilities of Large Language Models