We can group the conclusions based on a level of compression and the probe sets into two

parts: i) higher compression rates (0.5, 0.3 and in some cases even 0.2 bpp) seem to be

suitable for recognizing faces with different expressions (

fb probe set) and images taken in

different illumination conditions (fc probe set) ii) lower compression rates (1 bpp) seem to

be suitable for recognizing images taken at different points in time (

dup1 and dup2 probe

set). Taking this analysis into account, it seems that the current practice of deciding on the

level of compression based on visual distortion of images is wrong. While the images

compressed to 0.3 bpp are visually significantly distorted, the recognition results are in

almost all experiments statistically indistinguishable from the results achieved by using

uncompressed images. In many cases these results are slightly better and in some cases even

significantly better than the ones achieved with uncompressed images.-We can group the conclusions based on a level of compression and the probe sets into two

parts: i) higher compression rates (0.5, 0.3 and in some cases even 0.2 bpp) seem to be

suitable for recognizing faces with different expressions (

fb probe set) and images taken in

different illumination conditions (fc probe set)　ii) lower compression rates (1 bpp) seem to

be suitable for recognizing images taken at different points in time (

dup1 and dup2 probe

set). Taking this analysis into account, it seems that the current practice of deciding on the

level of compression based on visual distortion of images is wrong. While the images

compressed to 0.3 bpp are visually significantly distorted, the recognition results are in

almost all experiments statistically indistinguishable from the results achieved by using

uncompressed images. In many cases these results are slightly better and in some cases even

significantly better than the ones achieved with uncompressed images.