Multi-UAV system enhances map creation with increased adaptability and broader coverage. However, current techniques for deploying multiple UAVs encounter difficulties such as precision in collaborative localization, scale variance, and low processing speed for mapping. To tackle aforementioned challenges, a novel cluster-based framework designed for real-time collaborative localization and mapping is proposed. The framework adopt a well designed visual odometry system for handle large-scale scene. UAVs achieve collaborative localization through a dual-stage optimization process that includes a joint optimization algorithm and a collaborative localization refinement method, which ensures precise UAV localization and reduces scale variance. These estimated positions are then utilized for the global and dense mapping. To assess the effectiveness of proposed approach, both qualitative and quantitative tests using real-world datasets are conducted. The findings indicate that our framework is capable of significantly reducing scale variance and producing large-scale, consistent dense mapping.