Effective user training could help us to improve the discrimination performance of our intention in brain computer interface (BCI). This paper aims to differentiate users left or right hand motor imagery (MI) tasks with different scenarios in 3D virtual environment, as non-object-directed (NOD) scenario, static-object-directed (SOD) scenario and dynamicobject-directed (DOD) scenario respectively. The results have significant differences by applying these three scenarios. Both SOD and DOD scenarios pro-vide better classification accuracy, shorten single-trial period, and need smaller training samples comparing with the NOD case. We conclude that improving visual display may facilitate learning to use a BCI. Further comparing these results between single-subject and multiple-subject paradigm of BCI, we verify better classification performance could also be achieved by the multiple-subject paradigm. We believe these findings have the potential to improve discrimination performance of users intention for EEG-based BCI applications.