The visibility of streamed video distortions is affected by the user's display and viewing conditions, such as their display's luminance or the viewing distance. Distortions become less visible at lower luminance levels or farther viewing distances. However, most image and video quality metrics assume fixed optimal display and viewing conditions, making them insensitive to the viewing conditions. In this work, we demonstrate how existing metrics can be adapted to account for both factors and compare such an adaptation with the metrics that natively model viewing conditions. First, we collected a new video quality dataset, HDR-VDC, which contains streaming distortions in 4K HDR videos. The dataset consists of 12 4K and 4 FHD HDR source video content, each distorted with AV1 encoding and Lanczos upscaling from a lower resolution. The quality loss due to the distortions was measured at two viewing distances, corresponding to 60 and 120 pixels per visual degree, and two display mean luminance levels, varying by a factor of 8 in terms of luminance. Next, we developed a pre-processing framework that lets us adapt many quality metrics so that they can evaluate content across the viewing distances and luminance levels. The framework employs a display model and perceptually uniform encoding, PU21, to account for the effect of display luminance, and perspective projection scaling (PPS) to account for the effect of viewing distance. We used the framework to compare the performance of 35 state-of-the-art metrics. This allowed us to identify a selection of existing quality metrics that can be effectively adapted to account for the viewing conditions.
Figures & tables
Figure 1. The visibility of streaming distortions depends on viewing conditions. The same compressed 4K HDR content exhibits fewer visible artifacts at lower display luminance or greater viewing distances. We introduce HDR-VDC, a dataset that captures these effects across luminance levels and viewing distances, and evaluate whether existing quality metrics can be adapted using a display model, perceptual encoding, and perspective projection scaling. Our results show that, while some metrics can account for these factors, many remain unreliable, underscoring the need for quality metrics that natively model real-world viewing conditions.
Figure 2. Screenshots of the selected reference videos (The sequences are tone-mapped with ffmpeg using the ‘hable’ algorithm ( Hable, 2010 ) to BT.709 for visualization). For each selected video, we show its resolution and framerate.
Figure 3. Characterization of the selected reference videos. Indices correspond to the video indices in Figure 2 . Differences between SI/TI and complexity measures arise from their distinct formulations: SI/TI operate on pixel-level gradients and frame differences, whereas complexity measures block-wise DCT texture energy.
Figure 4. The configurations corresponding to the combination of two viewing distances: Near-Near, Far-Far, Far-Near, and Near-Far (we show one of the two cross-viewing distance settings in this figure).
Figure 5. The effect of viewing distance and display luminance level on the perception of quality for different resolutions and bitrates (compression levels) and across all contents. The bitrate reported in the figures is the average bitrate across all contents. The error bars represent the 95% confidence intervals.
Figure 6. Histograms of the JOD scores and their distribution (estimated using the kernel density estimate for visualization purposes) across viewing distances and display luminance levels. A few conditions exceed the score of 10 JOD due to the measurement noise.
Figure 7. Quality-bitrate curves for the RushPixar content (index 10 in Figure 2 ) for each viewing distance and luminance level. The shaded areas represent 95% confidence intervals. Dashed lines mark the approximate bitrate needed to reach 9.5 JOD. A few conditions exceed the score of 10 JOD due to measurement noise.
Figure 8. An illustration of the pipeline used to simulate the human visual system perception of HDR videos across viewing distances and display luminance levels, prior to the metrics evaluation. The display model component is employed for all metrics. However, metrics that natively account for the viewing distance do not employ PPS rescaling, and photometric metrics that utilize linear RGB input do not use the PU21 encoding.
Figure 9. (a) Display model response and (b) comparison of encoding functions.
Figure 10. SROCC performance of the metrics when accounting for the viewing distance and/or the display luminance level (by employing the PPS rescaling and the luminance level multiplier presented in Figure 8 ) and when not accounting for the viewing distance and display luminance (no PPS rescaling and luminance level multiplier). Error bars denote 95% confidence intervals. The * or ** symbols indicate a statistically significant difference in performance from the baseline (not accounting for the viewing conditions) at p<0.05 and p<0.01 (one-tailed paired t-test), respectively.
Figure 11. SROCC performance indices of the metrics on the HDR-VDC dataset. The metrics are ordered from the best (top-left) to the worst (bottom-right). Non-photometric quality metrics and/or metrics that do not natively account for viewing distance were adapted using PU21 encoding and/or PPS rescaling (as indicated in the label). The violin plots illustrate the distribution of the means that model variability due to content selection and observer responses. The black “+” denotes the mean score, and the red and blue “-” denote the 2.5th and 97.5th percentiles, which estimate the 95% confidence interval. The solid lines group the metrics for which there is no statistical significance in the difference in their performance.
Figure 12. The Reduced Chi-Square Xv2 performance (the lower, the better) of the disentanglement analysis of both the effect of viewing distance and the display luminance level.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
Video Content
HDR Content
Compression Distortions
Upscaling Distortions
Account for Viewing Distance
Account for Display Luminance
Publicly Available
Test Methodology
Liu et al. (2014)
No
No
Yes
No
Yes
No
Yes
ACR
Amirpour et al. (2021)
Yes
No
Yes
No
Yes
No
Yes
ACR
Sugito et al. (2022)
Yes
No
Yes
Yes
Yes
No
No
ACR
Kufa and Kratochvil (2019)
Yes
No
Yes
Yes
Yes
No
Yes
PC
Čadík et al. (2011)
Yes
Yes
Yes
No
No
No
Yes
DRM
Azimi et al. (2014)
Yes
Yes
Yes
No
No
No
No
DSIS
Appendix
Table 1. A comparison between our proposed HDR-VDC dataset and existing datasets in the literature. In the “Test Methodology” column, “ACR” refers to the absolute category rating methodology, “PC” to the pairwise comparison protocol, “DRM” to distortion regions marking, “DSIS” to double stimulus impairment scale, and “4AFC” to the four-alternative forced-choice.
Source sequence index
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
High-quality encoding CRF
7
9
9
5
9
9
9
9
7
7
11
7
9
7
7
7
Medium-quality encoding CRF
43
37
43
33
35
41
49
37
43
33
45
27
37
37
36
55
Low-quality encoding CRF
61
58
62
62
53
59
62
60
60
51
59
55
58
56
52
62
Appendix
Table 2. CRF settings of the distorted videos. The same CRF value was used across all resolutions. The reference videos are the high-quality encoded videos (at their native resolution).
Figure 13. Per-content effect size ( η2 ) of the viewing distance (top) and the display luminance level (bottom) as a function of the content characteristics of the 16 reference videos, with the Pearson correlation coefficient ( r ), its p -value, and a least-squares fit (dashed line); Source indices correspond to the video indices in Figure 2 .
Dataset
Content Type
Test Methodology
Viewing Distance [ ppd ]
Viewing Distance Effect Size
Display Luminance [ cd/m 2 ]
Display Luminance Effect Size
CID:IQ ( Liu et al., 2014 )
SDR Images
ACR
31.5 63
0.035
-
-
VCIP21 ( Amirpour et al., 2021 )
SDR Videos
ACR
60 92 128
0.063
-
-
VLIC ( Mikhailiuk et al., 2021b )
SDR Images
4AFC & QUEST
30 60
0.109
220 10
0.305
HDR-VDC
HDR Videos
Pairwise Comparison
60 120
0.101
700 600
0.015
Appendix
Table 3. A summary of the datasets used in the comparison of the effect size of the viewing distance and display luminance on the visibility of distortions.
Figure 14. Representative frames of the 16 reference videos as encoded (Original) and as simulated on the LG OLED G2 in the bright and dim conditions, each with the luminance histogram of the whole video, its dynamic range (DR, blue band), mean luminance (orange) and peak luminance (red).
Figure 15. JOD scores versus FovVideoVDP, HDR-VDP-3, FSIMc, and FUNQUE predictions (columns). Rows 1–2: near and far viewing distance (same test videos connected) for the bright and dim display; rows 3–4: disentangled scores and predictions (far − near). Rows 5–8: the same for display luminance (dim − bright) at near and far viewing distance.
Video quality assessment (VQA) plays a critical role in optimizing video delivery systems. While numerous objective metrics have been proposed to approximate human perception, the perceived quality strongly depends on viewing conditions and display characteristics. Factors such as ambient lighting, display brightness, and resolution significantly influence the visibility of distortions. In this work, we address the question of the multi-screen quality assessment on mobile devices, as this area still tends to be under-covered. We introduce a first large-scale subjective dataset collected across more than different 300 Android devices, accompanied by metadata on viewing conditions and display properties. We propose a strategy for aggregated score extraction and adaptation of VQA models to device-specific quality estimation. Our results demonstrate that incorporating device and context information enables more accurate and flexible quality prediction, offering new opportunities for fine-grained optimization in streaming services. Ultimately, this work advances the development of perceptual quality models that bridge the gap between laboratory evaluations and the diverse conditions of real-world media consumption. We made the dataset and the code available at https://videoprocessing.github.io/device-viewing-conditions.
Nikolay Safonov, Dmitriy S. Vatolin
MSU Institute for Artificial Intelligence, Moscow, Russia · AI Center, Lomonosov Moscow State University, Moscow, Russia
Avatar-streaming systems are commonly evaluated with image and video quality assessment (IQA/VQA) metrics, implicitly treating visual fidelity as a proxy for communicative success. We test this assumption through a controlled behavioral study of rendered 3D avatars across a pristine condition and fourteen geometric, photometric, temporal, and combined distortions. Fifty-nine participants contributed 2,688 judgments of perceived action, response confidence, and visual quality. We identify Misleading Quality in this dataset as distorted renderings that retain above-average perceived quality but yield below-average action-recognition accuracy. We also derive an Intent Quality Score (IQS) combining recognition correctness and confidence as the behavioral target for objective metrics. Among 126 distorted content--condition cells, 31 (24.6%) exhibited Misleading Quality; temporal and geometric distortions showed the highest rates, at 50.0% and 31.1%, respectively. The results reveal a quality--accuracy dissociation where distortion families affect appearance and communication differently. Across 24 direct-scoring IQA/VQA metrics and three supervised feature-regression baselines, alignment with IQS remained limited; at λ=0.5, the best leave-one-content-out baseline reached PLCC =0.4435. Under this controlled protocol, visual fidelity alone is insufficient for avatar communication, motivating intent-aware quality assessment and streaming objectives.
Ning-Hsuan Chang, Kai-Siang Ma, Yu-Chih Chen
National Chengchi University Taiwan · National Yang Ming Chiao Tung University Taiwan
Video technology is advancing toward Ultra High Definition (UHD) and High Dynamic Range (HDR), which intensifies the need for higher compression efficiency for these high-specification videos. Beyond advances in traditional codecs, neural video codecs (NVCs) have attracted significant research attention and have evolved rapidly over the past few years. The coding artifacts of NVCs often exhibit content-varying and generative characteristics, which differ from those of conventional codecs and are challenging for traditional video quality assessment (VQA) methods to capture. Therefore, VQA metrics are required to generalize across different codecs, content types, and dynamic ranges to better support video codec research and evaluation. In this paper, we propose FDIM, a feature-distance-based generic video quality metric for both traditional and neural video codecs across SDR and HDR formats. FDIM employs a hybrid architecture that integrates deep and hand-crafted features. The deep feature component learns multi-scale representations to capture distortions ranging from structural and textural fidelity degradation to high-level semantic deviations, while the hand-crafted feature component provides stable complementary cues to improve overall generalization. We trained FDIM on a large-scale subjective quality assessment dataset (DCVQA) consisting of over 16k video sequences encoded by traditional block-based hybrid video codecs and end-to-end perceptually optimized neural video codecs. Extensive experiments on ten SDR/HDR VQA datasets containing diverse, previously unseen codecs demonstrate that FDIM achieves strong generalization and high correlation with subjective assessment. The source code for FDIM and the DCVQA validation set will be released at https://github.com/MCL-ZJU/FDIM.
Jiayi Wang, Lichun Zhang, Xiaoqi Zhuang +3
College of Information Science and Electronic Engineering, Zhejiang University, and also with the Zhejiang Key Laboratory of Multimodal Communication Networks and Intelligent Information Processing, Hangzhou 310058, China · Central Media Technology Institute, Huawei, Hangzhou 310058, China