The Counting Bloom Filter (CBF) is a space-efficient data structure that extends a Bloom filter so as to allow approximate multiplicity queries on a dynamic multi-set. An in-depth study of three existing CBF schemes is presented, that is, the Na?ve Counting Bloom Filter (NCBF), the Space-Code Bloom Filter (SCBF) and the d-left Counting Bloom Filter (dlCBF). Then, a CBF scheme called Binary Shrinking d-left Counting Bloom Filter (BSdlCBF) is proposed. A performance metrics named load adaptability for CBF schemes is also defined. The performance of the four CBF schemes is evaluated by using metrics of counting error, space complexity and load adaptability under both uniform and Zipfian multiplicity distributions. The experimental results show that the proposed BSdlCBF outperforms the other three in terms of accuracy, space-efficiency and load adaptability. The cost of such an advantage of BSdlCBF is a reasonable rise in computational and space complexity.