SVG Tech Insight: Bit-Rate Evaluation of Compressed HDR Using SL HDR1
This fall SVG will be presenting a series of White Papers covering the latest advancements and trends in sports-production technology. The full series of SVG’s Tech Insight White Papers can be found in the SVG Fall SportsTech Journal HERE.
A number of current HDR standards include transmission of metadata along with the content. HDR encodes absolute luminance information, which may be outside the limits of what a particular monitor can display. Metadata helps the monitor map the incoming content to its capabilities.
The SL‑HDR1 standard  operates this way. What is unique about it is the fact that the content is actually SDR, and the metadata allows a compatible device to map that SDR content to the original HDR, or to some intermediate level that it can support. Legacy devices with no metadata support can simply display SDR. This is similar to what happened when color TV was introduced — the transmission was a black-and-white image with color information on the side.
Video encoders and decoders are agnostic to HDR. The encoder takes the video samples and converts them to a bit stream; the decoder converts that bit stream back to video samples that are approximately the same as what entered the encoder. Neither device interprets the meaning of the video samples. Metadata, if present, may be carried along the bit stream.
In this article, we attempt to answer the question of whether or not there is a bit rate penalty when one uses SL‑HDR1 to transport HDR over a compressed video link. We do that by establishing a baseline with native HDR video, and then switch to SL‑HDR1 and determine at what bit rate the quality is the same as the native HDR. This is similar to the work presented in .
Bit Rate Evaluation
The test setup is shown in Figure 1. As indicated, there are three test paths:
- Path 1 (in blue) is an end-to-end HDR10 path. A native SDI HDR10 signal is applied directly to the encoder, converted to either AVC or HEVC, and then decoded back to SDI.
- Path 2 (in purple) is an SL‑HDR1 path. The native SDI HDR10 signal is routed to an SL‑HDR1 encoder, which produces an SDI SDR signal with metadata, carried in the ancillary data space using SMPTE-2108 . This signal is applied to the encoder and converted to either AVC or HEVC. The SL‑HDR1 metadata is extracted from the ancillary space and injected in the video bit stream as SEI messages. The decoder produces an SDI signal, and with the metadata restored to the ancillary space. Finally, an SL‑HDR1 decoder re-creates the SDI HDR10 signal.
- Path 3 (in red) is an SL‑HDR1 path that bypasses the encoder/decoder. It is used to obtain a baseline reading without compression.
In all cases, both the original and decoded signals are captured in YUV format by a video recorder. The files are transferred to a computer, where the quality metrics are calculated. For this evaluation, we selected the following metrics:
- Peak Signal-to-Noise Ratio (PSNR), which measures the absolute difference between each frame in a sequence. It is well-known that PSNR does not correlate well with perceived quality.
- PSNR_DE100: PSNR of mean of absolute deltaE2000 metric, referred to a 100‑nit luminance.
- PSNR_L100: PSNR of mean square error of L component of CIELab color space used for the deltaE2000 metric, referred to a 100‑nit luminance.
The DE100 and L100 metrics were selected since they have been shown to correlate well with perceived quality .
The test procedure was as follows:
- Take a baseline reading of the PSNR using Path 3. This only needs to be done once.
- Select a target test video bit rate Br for the AVC/HEVC encoder.
- Run the Path 1 signal and compute the selected quality metrics.
- Run the Path 2 signal and compute the selected quality metrics.
- Repeat steps 2-4 for other values of Br.
- Perform the BD-rate computation for both PSNR_DE100 and PSNR_L100.
After the PSNR results were obtained, we performed the BD-rate computation  in order to evaluate the average increase or decrease in bit rate in the SL‑HDR1 case to match the quality metric in the HDR10 case. This process was done only for the metrics that correlate well with perceived quality, namely PSNR_DE100 and PSNR_L100.
The details of the test setup in Figure 1 are:
- SL‑HDR1 Encoder and Decoder: Cobalt 9904-UDX
- AVC/HEVC Encoder: Cobalt 9992-ENC
- GOP size: 100 frames
- Bit Depth: 10 bits
- Chroma Mode: 4:2:0 (consumer grade signals)
- AVC/HEVC Decoder: Cobalt 9992-DEC
The tests were performed with three test sequences. All sequences had the following common parameters:
- Duration: 12 seconds
- Resolution: 1920×1080
- Color Space: BT 2020
The contents of each sequence were as follows:
- Sequence 1: “base jump” — extreme sports in mountains
- Sequence 2: “baseball” — baseball game at night
- Sequence 3: “zombie” — city scenes
Table 1 presents the quality metrics of the SL‑HDR1 process before encoding and decoding (Path 3 in Figure 1). The SL‑HDR1 process is not exact – there is a small impact in the metrics, as the image is not exactly reconstructed. Note that the YUV PSNR is provided as reference; since these metrics are fundamentally different, the absolute values should not be compared between them.
Table 1: SL‑HDR1 Metrics with Encoder/Decoder Bypassed
Figure 2 shows the raw PSNR_DE100 and PSNR_L100 results for the three sequences, using both HEVC and AVC encoding. The bit rate ranges used are different since HEVC encoding is more efficient than AVC encoding, so a higher range is used for AVC.
Generally, Figure 2 shows that there is almost always some PSNR improvement at the same bit rate when SL‑HDR1 is used, with very few exceptions. In other words, even though the SL‑HDR1 process is not perfect (the HDR image is exactly reproduced), when one uses quality metrics that correlate well with perceived quality, the resulting image after compression/decompression actually looks better than compressing and decompression the HDR image directly. This matches the conclusions presented in .
The remaining question pertains to quantifying the bit rate advantage. If one is seeking to achieve a certain target quality for an HDR link and has the option of either transmitting HDR10 natively or SL‑HDR1, which method will yield the lowest bit rate and by how much? The standard way of answering this question is by the use of BD-rate , which produces an “average” value over the tested range. The relevant numbers are provided in Table 2 below. In this table, positive values indicate that the SL‑HDR1 bit rate for the same quality is higher, and negative values indicate that it is lower.
Table 2: BD-Rate Values
Table 2 indicates that, for most combinations, a lower bit rate is required for SL‑HDR1 transport as compared with a straight HDR10 link, sometimes significantly so. This confirms the conclusions presented in , using a different commercial encoder.
One final item for discussion pertains to the SL‑HDR1 metadata bit rate. When transporting HDR10, there is no metadata requirement, but SL‑HDR1 adds per-frame metadata that is included in SEI messages inside the video elementary stream. The maximum amount of metadata per frame is 61 bytes. Therefore, an upper bound on the SL‑HDR1 metadata bit rate is 24.4 kb/s for a 50‑fps signal, and 29.3 kb/s for a 60‑fps signal. This bit rate increase is negligible (less than one audio channel), and, in many encoders, is absorbed by a slight adjustment in the NULL packet rate, so the overall bit rate is unchanged.
The impact of SL-HDR1 in compressed bit streams is a function of the content, and can be quite significant. When using quality metrics that are better correlated with human perception, it is often possible to actually decrease the link bit rate while keeping the same quality, as originally reported in .
- European Telecommunications Standards Institute, “High-Performance Single Layer High Dynamic Range (HDR) System for use in Consumer Electronics devices; Part 1: Directly Standard Dynamic Range (SDR) Compatible HDR System (SL‑HDR1),” ETSI TS 103 433-1 V1.2.1, 2017
- Touze, D., and Kerkhof, L., “Single-Layer HDR Video Coding with SDR Backward Compatibility”, 2017 SCTE-ISBE CABLE-TEC EXPO
- Society of Motion Picture and Television Engineers, “HDR/WCG Metadata Packing and Signaling in the Vertical Ancillary Data Space,” SMPTE ST 2108-1, 2018
- Hanhart, P., Řeřábek, M., and Ebrahimi, T., “Towards high dynamic range extensions of HEVC: subjective evaluation of potential coding technologies”, SPIE 9599, Applications of Digital Image Processing XXXVIII, 95990G (22 September 2015)
- Bjontegaard, G., “Improvements of the BD-PSNR model”, ITU-T SG16/Q6 VCEG 35th meeting, Berlin, Germany, 16–18 July, 2008, Doc. VCEG-AI11