We began by studying the papers listed in the References section below.
It was noted that Motion Compensated Frame Interplation (MCFI) was preferred
over Optical Flow and Pel-Recursive methods due to its lower computational
requirements and decent performance. Three MCFI methodologies were implemented
in VC++ on the Windows NT platform. It was found that MCFI using Bidirectional
Motion Estimation [1] and MCFI using using a pixel classifier [2] required roughly
the same computational power and yielded very similar results. The pixel classification
method was found to yeild a slightly higher PSNR for the Mobile and Akiyo sequences, but a
lower PSNR for Stefan. The pixel classification method should work quite well for talking
head sequences.
In addition to implementing the two methods mentioned above, a MCFI alorithm that utilized only
backward motion vectors was implemented. (as opposed to using just both forward and backward motion
vectors in the other two methods). This method can take advantage of the backward motion vectors
already coded in MPEG video sequences, so motion estimation calculations do not have to be redone.
Unfortunately, the PSNR of this method is quite low and many artifacts are produces, so this method
is not practical.