The JVT video coding standard (formerly H.26L), employs motion compensation with a 6-tap interpolation filter and variable block sizes. These are among the key features for realizing high compression ratio. However, they are also accompanied by an increase in the computational complexity and number of memory accesses-factors that are critical for real-time implementation of decoders especially for applications with larger frame size such as SDTV or HDTV. In this paper we show that an adaptive "switching" among a bank of interpolation filters can solve this problem. The proposed approximation method for motion estimation at the encoder requires no additional frame memory for switching between interpolation filters. We also present an analysis of the encoder complexity for the proposed method. Simulation results show that the proposed method significantly reduces the number of memory accesses without sacrificing coding efficiency and subjective quality.