H.261/nv Compression Performance

Excerpted from our ACM Multimedia paper.

rate-distortion comparison

Relative compression performance of nv and Intra-H.261.

Because the Intra-H.261 and nv compression schemes use similar conditional replenishment algorithms, we can evaluate their relative compression performance simply by ignoring the temporal dimension and comparing only their 2D image compression performance. The figure above shows the performance of the two approaches for the canonical 8-bit, 512x512 grayscale ``Lena'' image. Both encoders were modified to omit block-addressing codes to allow the H.261 encoder to operate on a non-standard image size. This modification has little impact on the results since block-addressing accounts for a small fraction of the bit rate.

The peak signal-to-noise ratio (PSNR) is plotted against rate (in bits per pixel). Multiple points were obtained by varying the H.261 scalar quantizer and the nv dead-zone threshold. Note that the nv algorithm was intended to operate with a non-configurable, fixed threshold, but we explored other thresholds to complete a rate-distortion curve.

As seen in the graph, H.261 consistently outperforms the nv coder by 6-7dB. Since transmissions are typically rate-limited, we should consider a fixed distortion and compare the corresponding bit rates. From this perspective, the nv bit rate is two to three times that of H.261. For a rate-limited transmission, this translates into a factor of two to three decrease in frame rate.