In recent years, how to achieve stable localization and construct high-quality dense maps in large-scale scenes has become a research highlight. In large-scale scenes, for the consideration of the mapping accuracy and efficiency, multi-agent systems rather than single-agent ones are usually employed. Currently, as far as we know, collaborative VI-SLAM (Visual Inertial Simultaneous Localization And Mapping) systems applicable to multi-agent systems are still sporadic, and systems those can achieve a good balance among the localization accuracy, the mapping density, and the transmission efficiency are temporarily lacking. In this paper, we propose a novel centralized collaborative VI-SLAM framework, namely TES-CVIDS (Transmission Efficient Sub-map based Collaborative Visual-Inertial Dense SLAM). In TES-CVIDS, instead of the original RGBD images, the compact sub-maps are transmitted, effectively reducing the transmission data redundancy. After that, the server completes key-frame processing, hierarchical pose-graph optimization, and global dense map construction in three separate threads. Besides, thanks to our depth search mechanism, the geometry information of all key-frames can be recovered on the server-end. Thus, sub-maps can be regenerated after the global pose-graph optimization to maintain the consistency between the localization and the mapping. Both the qualitative and the quantitative experimental results corroborate the superior performance of our TES-CVIDS. To make our results reproducible, the source code has been released at https://cslinzhang.github.io/TES-CVIDS-MainPage/.