This contribution proposes the reduction of the worst-case memory bandwidth and operation number of overlapped block motion compensation (OBMC) prediction which is related to Core Experiment 10.2 (CE10.2). The worst-case memory bandwidth and operation cost of OBMC are studied in CE10.2 and applying uni-prediction constraint on OBMC were recommended to reduce them in the Macao meeting. The memory bandwidth and operation number of bi-directional MC prediction was placed as the target criteria of the reduction but that of CE10.2.1 is still over than the thresholds. In this contribution, two method are proposed to reduce memory bandwidth and operation furthermore; 1) modifying the OBMC filter according to CU size, 2) modifying uni-prediction constraint on neighbor block according to CU size. The two simulations are conducted under different OBMC applicable conditions which include the above modifications without exceedance of the worst-case memory bandwidth and operation number than bi-directional MC. In condition A, the OBMC filter is modified from 8-tap to 4-tap in all CU size and the uni-prediction constraint on neighbor block is removed from all CU size. In condition B, the OBMC filter tap and the constraint are changed depending on CU size. The condition A results show the BD-rate saving for luma by -0.37% in RA and -0.35% in LD, and the condition B results under condition B show the BD-rate saving for luma by -0.42% in RA and -0.43% in LD. It is confirmed that the proposed methods have the effect for not only reducing of worst-case memory bandwidth and operation number but also for improving the coding gains up to -0.15% in RA compared to that of CE10.2.1.