The technique of 1/8-pel interpolation [AD09] was proposed for motion-compensated prediction (MCP) and adopted in KTA software. Three types of interpolation filters are used for 1/2-, 1/4-, and 1/8-pel sub positions, respectively.
- [-3, 12, -39, 158, 158, -39, 12, -3]/256 for 1/2-pel sub positions.
- [-3, 12, -37, 229, 71, -21, 6, -1]/256 and [-1, 6, -21, 71, 229, -37, 12, -3]/256 for 1/4-pel sub positions.
- Bilinear filter for 1/8-pel sub positions.
The frequency response of the interpolation filter is shown in the following figure. As can be seen, it is almost an ideal low-pass filter with a gain of 8 and a cutoff frequency π/8.
According to the performance reported in the proposal, the gain on CIF/QCIF sequences is quite significant, i.e., up to 14% bit-rate reduction. I tested this technique based on a set of HD sequences. As shown in Table 1, the R-D performance is measured by BDPSNR [1], i.e., PSNR improvement at the same bit-rate or bit-rate reduction at the same PSNR.
For the 720p sequences (see the left half of Table 1), the PSNR gain varies from -0.24 dB to 0.11 dB, and the average is about 0.04 dB loss. The performance becomes even worse for 1080p sequences (see the right half of Table 1), where the average loss is 0.17 dB (about 4.83% bit-rate increasing) and the worst case is 0.58 dB loss. This observation shows that 1/8-pel MCP is not an efficient tool for HD video coding. As an aliasing-compensated coding tool, it is supposed to perform better on low resolution sequences, where aliasing is more likely to occur.
[1] G. Bjontegaard, ”Calculation of average PSNR differences between RD-curves,” in ITU-T SG16/Q6, VCEG-M33, Apr. 2001.
Permanent Link: R-D Performance of 1/8-pel MCP on HD Sequences
# 2009-06-08 Monday 4:30 pm
Why measure quality with PSNR? That metric blows.
# 2009-06-08 Monday 7:29 pm
Yes. PSNR has some problems. But in the framework of hybrid video coding, where distortion is caused by quantization in the transform domain, PSNR correlates well with the perceived quality, although not strictly increasing with it.
The experiment herein follows the test condition and the evaluation criterion specified by VCEG. Now, VCEG is considering using VQM to evaluate the subjective quality. But I think PSNR will be relevant in quality measurement for quite a long time.
# 2009-06-12 Friday 4:12 pm
Optimizing an algorithm for PSNR will deprive output quality of what it could’ve had if it was optimized for something like SSIM. I haven’t checked out VQM but it’s cool as long as they take the step up from PSNR. It is NOT consistent for perceived quality in my experience.
# 2009-06-25 Thursday 12:02 pm
Images have order-3 patterns. A better way of doing interpolation at 1/8-pixel sub positions would be to fit a 3rd-order polynomial through the 4 neighboring pixels on the same line, then use the polynomial to interpolate.
# 2009-09-14 Monday 11:40 am
Your result is compared to the “anchor” configuration, right?
I’m curious that why there is negative effect of using 1/8-pel MCP compared to 1/4-pel, since 1/8-pel MCP just search more sub-positions. I think the result should be “same or better”, no matter how bad are the 1/8-pel sub-positions interpolated.
Any ideas or hints?
# 2009-09-14 Monday 3:42 pm
MV’s resolution is increased, which means that more bits are needed to represent the same displacement, compared with 1/4-pixel MCP.
# 2009-09-14 Monday 5:18 pm
Many thanks for your reply.
Yes, more bits are required to represent the higher precision MVs, but is that possible to have a situation like the “Station” sequence that having a 16.15% bitrate increase? Since it is only one bit per motion vector.
I read some comment from the Internet that suggest that it may be due to the mismatch of SATD(or SAD) used in MV search with the SSD used in RDO, how do you think about that?