When standard coded video is transmitted over Internet where packets may be lost, there exists the annoying problem of error propagation. In this paper we propose a new video coding scheme that dynamically chooses multiple references from multiple candidate frames in prediction, i.e., dynamic multi- reference prediction (DMRP). Based on the system model of the video codec with DMRP, we derive that the more candidate frames from which the references are chosen, the greater error resilience is achieved. Specifically we study the performance of video coding with two references selected from four candidate frames (2R4CF). The simulation results under the typical packet loss rates of Internet show that video coding with 2R4CF provides better error resilience than fixed two-reference video coding [4], single-reference four-frame video coding [3] and H.263 standard coding [5] with the coding efficiency nearly the same or a little higher than H.263 standard coding.