CRF vs CBR vs QP Dataset

Compare the CRF, CBR and QP modes.

The parameters of this database

Name	Value	Description (link)	Total
effort	fast, medium	enc_effort	\(2\)
encoder	libsvtav1, libx264, libx265	enc_encoder	\(2\)
host	paradoxe-26	env_hostname	\(1\)
metrics	psnr, ssim	Table t_met_metric
mode	crf, cbr, qp	enc_mode	\(3\)
profile	sd, fhd	video resolution and more	\(2\)
quality	\(\left\{\frac{k}{17} \mid k \in [1,16] \right\}\)	enc_quality	\(16\)
repeat	\(5\) times	total repetitions	\(5\)
video	baolei_balloon, park_joy	ctc	\(2\)
total	\(\prod\)	total number of points	\(5760\)

Download

mendevi download crf_vs_cbr_vs_qp.db.xz.torrent

• crf_vs_cbr_vs_qp.json.xz (914 Ko).

• crf_vs_cbr_vs_qp.db.xz.torrent (593 Mo).

Plots

Effect of mode on energy

The graph below shows the energy impact of the QP for a comparable bitrate.

mendevi plot crf_vs_cbr_vs_qp.db -x bitrate -y energy -wx profile -wy encoder -c mode -m effort -e quality -f 'ref_stem=="park_joy"'

The graph below shows the energy impact of the QP for a comparable quality.

mendevi plot crf_vs_cbr_vs_qp.db -x psnr -y energy -wx profile -wy encoder -c mode -m effort -e quality -f 'ref_stem=="park_joy"' -t t_enc_encode

Conclusion

On this machine and in this video, the encoder, the profile and the force have a much greater impact on the encoding energy than the mode. The difference between qp and crf modes is imperceptible.

The cbr mode appears to be more efficient than the others for the libsvtav1 encoder.

Reproduce

Version 4.2.0 of libsvtav1 does not support optimisation on ssim and the random/access prediction on cbr mode. Although mendevi enables these default settings for the qp and crf modes, they need to be always disabled here to ensure a fair comparison between the different modes of libsvtav1. In order to optimise PSNR and predict with low delay, you have to create the file callback.py with the following content:

import re

from mendevi.cmd import CmdFFMPEG

def libsvtav1_lowdelay_psnr(cmd: CmdFFMPEG, **kwargs) -> CmdFFMPEG:
    """Set tune=1 and pred-struct=1 to the svtav1 parameters."""
    if kwargs["encoder"] != "libsvtav1":
        return cmd
    encode = cmd.encode
    try:
        idx = encode.index("-svtav1-params") + 1
    except ValueError:
        encode.append("-svtav1-params")
        idx = len(encode)
        encode.append("")
    encode[idx] = (
        re.sub(r"tune=\d", r"tune=1", encode[idx])
        if "tune=" in encode[idx]
        else f"{encode[idx]}:tune=1"
    )
    encode[idx] = (
        re.sub(r"pred-struct=\d", r"pred-struct=1", encode[idx])
        if "pred-struct=" in encode[idx]
        else f"{encode[idx]}:pred-struct=1"
    )
    cmd.encode = encode
    return cmd

for video in baolei_balloon.mkv park_joy.mkv; do
    mendevi prepare $video -p sd
    mendevi prepare $video -p fhd
done

mendevi encode -m crf -m cbr -m qp -r 5 -c libsvtav1 -c libx264 -c libx265 -e fast -e medium --callback callback.py reference*
mendevi decode sample*

mendevi probe reference* sample*