LIVE-YT-Banding Benchmark Analysis

• LIVE-YT-Banding is a video dataset with subjective ratings of banding severity in streaming content, serving as a benchmark for no-reference artifact detection and QoE analysis.
• It supports evaluation of methods like CAMBI, FS-BAND, and BBAND, which use contrast-aware multiscale processing, edge-preserving smoothing, and dual-CNN architectures to identify false contour artifacts.
• The dataset underscores challenges in correlating pixel-domain banding with perceived quality and highlights the need for temporal pooling, calibration, and integration with streaming-system QoE models.

LIVE-YT-Banding, or LIVE YouTube Banding, denotes a video dataset with subjective ratings of banding severity during YouTube-like streaming, and it functions as a natural evaluation target for no-reference banding predictors, perceptual visibility models, and streaming-oriented QoE analysis. In the literature summarized here, LIVE-YT-Banding is repeatedly treated as a relevant benchmark context rather than as the native training or validation corpus of the principal algorithms: CAMBI, FS-BAND, and BBAND all address the same broad phenomenon of quantization-induced false contours, but none of them directly uses LIVE-YT-Banding in the reported experiments (Tandon et al., 2021, Chen et al., 2023, Tu et al., 2020). A separate line of work on YouTube Live streaming, exemplified by ReCLive, addresses encrypted-traffic-based QoE inference rather than pixel-domain banding detection, yet it is operationally relevant because it characterizes the delivery conditions under which visible degradations become salient in live video (Madanapalli et al., 2021).

1. Scope and position within banding research

Within the cited literature, banding is defined as a false contouring artifact: artificially introduced, staircase-like contours that arise when a smoothly varying luminance or color field is quantized. The artifact is most conspicuous in spatially smooth regions with small gradients, such as skies, ocean, fog or dark scenes, and animation. The LIVE-YT-Banding setting is therefore most naturally situated at the intersection of perceptual video quality assessment, no-reference artifact detection, and streaming-system evaluation.

The papers that discuss methods relevant to LIVE-YT-Banding do so from different levels of analysis. CAMBI was designed for quantization-induced false contours in smooth regions and is presented as directly applicable to LIVE-YT-Banding with only minor adjustments to viewing-condition parameters and bit-depth handling (Tandon et al., 2021). FS-BAND is an image- and patch-level no-reference detector trained on DBI rather than on video data, so any use on LIVE-YT-Banding is explicitly described as an inference that must account for domain shift and the need for temporal pooling and calibration (Chen et al., 2023). BBAND is a no-reference video model for banding artifacts, but its reported validation is confined to the Wang et al. banding dataset rather than LIVE-YT-Banding (Tu et al., 2020).

Method	Uses LIVE-YT-Banding?	Reported relation
CAMBI	No	Designed for the same banding phenomena; described as directly applicable with viewing-condition adjustment
FS-BAND	No	Image/patch model that can be applied frame-by-frame; calibration and temporal pooling are needed
BBAND	No	Video-level banding predictor evaluated on Wang et al.’s dataset
ReCLive	No	Not a pixel-domain banding detector; infers live-stream QoE from encrypted traffic

A recurring source of confusion is that the term “banding” is used in two different senses across these materials. In CAMBI, FS-BAND, and BBAND, banding refers specifically to false contours created by quantization and compression. In the ReCLive deployment summary, the term is extended to bandwidth-driven live-stream degradations such as resolution drops and stalls. This suggests a broader operational usage in ISP-facing monitoring than in artifact-specific vision models (Madanapalli et al., 2021).

2. Perceptual and signal-theoretic basis

The common perceptual premise is that banding becomes visible when smooth gradients are converted into discrete plateaus separated by weak but extended edges. Lower bit-depth, coarser quantization, and certain codec settings increase the size and visibility of these steps. Larger displays and viewing geometries that enlarge plateau width in retinal terms also increase conspicuity. CAMBI formalizes this intuition by tying banding visibility to a contrast-aware, multiscale approximation of the Human Visual System’s Contrast Sensitivity Function, with contrast represented by the step magnitude $k$ and spatial-frequency dependence represented through visual degrees $v^\circ$ and the weight $\log_2(16/v^\circ)$ (Tandon et al., 2021).

Dithering plays a central role in this perceptual account. In CAMBI, undithered false contours in 8-bit content analyzed in a 10-bit space tend to appear as larger quantization steps, while dithered content shifts confidence toward smaller effective steps. The model therefore treats banding not merely as the presence of edges, but as the presence of visually meaningful plateau transitions after low-pass filtering and multiscale aggregation. This is important for LIVE-YT-Banding-style evaluation because YouTube-like delivery chains may involve bit-depth conversion, resolution changes, and codec choices that alter both the amplitude and spatial support of banding.

BBAND reaches related conclusions through a different HVS-inspired route. Its design is motivated by Mach bands and center-surround pooling in the early visual system, and it modulates edge visibility by luminance masking, texture masking, and edge cardinality rather than by an explicit CSF or JND threshold model (Tu et al., 2020). FS-BAND, by contrast, argues that banding is difficult to isolate in the spatial domain alone because subtle low-amplitude changes on smooth backgrounds can be confounded with genuine texture and edges; it instead emphasizes the complementary structure of high-frequency and low-frequency representations, without implementing an explicit CSF (Chen et al., 2023).

These viewpoints are compatible rather than contradictory. CAMBI emphasizes contrast step size and spatial scale; BBAND emphasizes visibility modulation of weak extended edges; FS-BAND emphasizes separability in frequency-sensitive representations. For LIVE-YT-Banding, this implies that the benchmark is best understood as testing both signal structure and perceptual visibility rather than simple edge presence.

3. Principal algorithmic families relevant to LIVE-YT-Banding

CAMBI is a no-reference, contrast-aware multiscale banding index operating on the luma channel. Its preprocessing consists of bit-depth normalization to 10-bit, a $2\times 2$ averaging low-pass filter to reduce dithering and emulate HVS low-pass behavior, and display normalization to 4k under the subjective-test assumptions. At each scale, it evaluates a contrast-aware confidence map

$$c(k,s)=p(0,s)\,\max\!\left[\frac{p(-k,s)}{p(0,s)+p(-k,s)},\,\frac{p(k,s)}{p(0,s)+p(k,s)}\right],\quad k\in\{1,2,3,4\},$$

where $p(k,s)$ is the fraction of low-gradient neighbors at offset $k$ within a fixed square window. It computes these maps on five downsampled resolutions—4k, 1080p, 540p, 270p, and 135p—corresponding to visual degrees $v^\circ\in\{1,2,4,8,16\}$. Frame-level pooling selects the worst $\kappa_p=30\%$ pixels and applies CSF-inspired weights:

$$\mathrm{CAMBI}_f=\frac{1}{|\kappa_p|}\sum_{(x,y)\in\kappa_p}\sum_{k=1}^{4}\sum_{v^\circ\in\{1,2,4,8,16\}} c\!\big(k,s(v^\circ)\big)\cdot k\cdot \log_2\!\left(\frac{16}{v^\circ}\right).$$

Temporal pooling samples frames every $v^\circ$0 and averages the resulting frame scores (Tandon et al., 2021).

BBAND is a no-reference video-level predictor centered on explicit banding edge extraction. It begins with self-guided, edge-preserving smoothing, computes Sobel gradients, and classifies pixels by thresholds $v^\circ$1 and $v^\circ$2 into flat pixels, texture pixels, and candidate banding pixels. Subsequent stages perform a uniformity check, non-maximum suppression, gap filling, edge linking, and edge-length filtering to obtain a banding edge map. Visibility is then estimated on the original frame using Gaussian-window local statistics, luminance masking, texture masking, and cardinality masking, integrated as

$v^\circ$3

Frame-level pooling averages the largest $v^\circ$4 percentile nonzero BVM values with SI weighting, and video-level pooling aggregates frames with TI weighting. BBAND is therefore explicitly spatial-temporal, although its temporal component is limited to weighting rather than motion-aware artifact modeling (Tu et al., 2020).

FS-BAND is a patch-based, no-reference detector that combines high-frequency and low-frequency representations in a dual-CNN architecture. The high-frequency map is obtained by Sobel gradient magnitude,

$v^\circ$5

while the low-frequency map is obtained through a piece-wise smooth approximation based on Mumford–Shah regularization. Two non-shared ResNet-50 branches process these maps separately, and features from the first convolutional layer and the last layer of each branch are concatenated and passed through fully connected layers to a 128-D vector and then to a sigmoid output for patch-wise banding probability. Visibility weighting is added through a spatial-frequency masking term based on local RMS-gradient activity, producing a pixel-wise banding map

$v^\circ$6

A global frame score is then computed by worst $v^\circ$7 pooling over the nonzero map values (Chen et al., 2023).

Taken together, these methods define three major methodological families for LIVE-YT-Banding-style analysis: multiscale CSF-informed contour confidence, HVS-masked banding-edge visibility, and learned dual-representation patch classification.

4. Experimental corpora and validation protocols

The CAMBI study was developed on a purpose-made Netflix banding dataset rather than on LIVE-YT-Banding. It used 9 short 4k, 10-bit source clips of duration 1–5 s, mostly exhibiting varying degrees of banding, with 1 clip having no banding. Each source was encoded at 1080p, quad-HD, or 4k; reduced to 8-bit using ffmpeg; and encoded with AV1 using libaom at $v^\circ$8. Ordered dithering was introduced during downsampling by ffmpeg and selectively pruned by the encoder depending on QP and resolution; a tenth encode per source at 4k, $v^\circ$9, was produced without dithering to explicitly study that condition. The subjective protocol used a modified Double Stimulus Continuous Quality Scale for worst-case banding annoyance with $\log_2(16/v^\circ)$0 and $\log_2(16/v^\circ)$1, 6 training anchors from an extra source, randomized exhaustive clip order, and 23 participants familiar with banding artifacts. A total of 86 encodes were scored after removing 4 high-quality AV1 encodes because of browser decode issues; all sequences had $\log_2(16/v^\circ)$2 and $\log_2(16/v^\circ)$3, and the 95% Student’s $\log_2(16/v^\circ)$4-confidence interval for MOS was below 10 (Tandon et al., 2021).

BBAND, by contrast, reports results only on the banding dataset created by Wang et al. The dataset consists of 6 clips of 720p videos at 30 fps, with VP9 encoding at different quantization levels to induce varying banding severity. MOS serves as the subjective ground truth, and a 5-parameter logistic mapping is fitted between predicted objective scores and MOS for PLCC and RMSE computation. LIVE-YT-Banding is explicitly stated not to be used for training, validation, or testing in this work (Tu et al., 2020).

FS-BAND uses the “Capturing banding in images” database released in Kapoor et al. (ICASSP 2021), often referred to as DBI. The paper does not report the exact number of images or patches, but it follows a random 8:2 train:test split on that database for patch-level banding classification. No video datasets are used, and no LIVE-YT-Banding links or identifiers appear in the study (Chen et al., 2023).

This experimental divergence is significant. LIVE-YT-Banding is positioned in these materials as a downstream benchmark or deployment target, while the core evidence for each method is obtained from separate, method-specific datasets with different distortion processes, granularity, and subjective protocols.

5. Reported performance and what it implies for LIVE-YT-Banding

On the Netflix banding dataset, CAMBI exhibits a strong, nearly linear, negative correlation with MOS: $\log_2(16/v^\circ)$5, $\log_2(16/v^\circ)$6, $\log_2(16/v^\circ)$7, and pairwise accuracy on statistically significant MOS differences $\log_2(16/v^\circ)$8. Among 3655 possible MOS comparisons for 86 videos, 2895 were statistically significant, and CAMBI correctly ordered 95% of those pairs. On a separate 4k, 10-bit HEVC dataset without visible banding comprising 84 encodes, CAMBI did not over-predict banding, and the paper reports the practical heuristic that $\log_2(16/v^\circ)$9 suggests “no visible banding.” The same study reports that VMAF and PSNR correlate poorly with banding MOS on that dataset (Tandon et al., 2021).

BBAND reports strong agreement with MOS on the Wang et al. dataset, outperforming two prior banding detectors. The reported values are $2\times 2$0, $2\times 2$1, $2\times 2$2, and $2\times 2$3, compared with Wang [11] at $2\times 2$4, $2\times 2$5, $2\times 2$6, $2\times 2$7, and Baugh [16] at $2\times 2$8, $2\times 2$9, $$c(k,s)=p(0,s)\,\max\!\left[\frac{p(-k,s)}{p(0,s)+p(-k,s)},\,\frac{p(k,s)}{p(0,s)+p(k,s)}\right],\quad k\in\{1,2,3,4\},$$0, $$c(k,s)=p(0,s)\,\max\!\left[\frac{p(-k,s)}{p(0,s)+p(-k,s)},\,\frac{p(k,s)}{p(0,s)+p(k,s)}\right],\quad k\in\{1,2,3,4\},$$1. No statistical significance tests are reported (Tu et al., 2020).

FS-BAND reports patch-level classification performance on DBI rather than video-level MOS correlation. Its results are $$c(k,s)=p(0,s)\,\max\!\left[\frac{p(-k,s)}{p(0,s)+p(-k,s)},\,\frac{p(k,s)}{p(0,s)+p(k,s)}\right],\quad k\in\{1,2,3,4\},$$2, $$c(k,s)=p(0,s)\,\max\!\left[\frac{p(-k,s)}{p(0,s)+p(-k,s)},\,\frac{p(k,s)}{p(0,s)+p(k,s)}\right],\quad k\in\{1,2,3,4\},$$3, accuracy $$c(k,s)=p(0,s)\,\max\!\left[\frac{p(-k,s)}{p(0,s)+p(-k,s)},\,\frac{p(k,s)}{p(0,s)+p(k,s)}\right],\quad k\in\{1,2,3,4\},$$4, and speed $$c(k,s)=p(0,s)\,\max\!\left[\frac{p(-k,s)}{p(0,s)+p(-k,s)},\,\frac{p(k,s)}{p(0,s)+p(k,s)}\right],\quad k\in\{1,2,3,4\},$$5. The paper compares against several baselines, including DBI (Kapoor et al. CNN) at $$c(k,s)=p(0,s)\,\max\!\left[\frac{p(-k,s)}{p(0,s)+p(-k,s)},\,\frac{p(k,s)}{p(0,s)+p(k,s)}\right],\quad k\in\{1,2,3,4\},$$6, $$c(k,s)=p(0,s)\,\max\!\left[\frac{p(-k,s)}{p(0,s)+p(-k,s)},\,\frac{p(k,s)}{p(0,s)+p(k,s)}\right],\quad k\in\{1,2,3,4\},$$7, accuracy $$c(k,s)=p(0,s)\,\max\!\left[\frac{p(-k,s)}{p(0,s)+p(-k,s)},\,\frac{p(k,s)}{p(0,s)+p(k,s)}\right],\quad k\in\{1,2,3,4\},$$8; HyperIQA at accuracy $$c(k,s)=p(0,s)\,\max\!\left[\frac{p(-k,s)}{p(0,s)+p(-k,s)},\,\frac{p(k,s)}{p(0,s)+p(k,s)}\right],\quad k\in\{1,2,3,4\},$$9; DBCNN at accuracy $p(k,s)$0; LPIPS at accuracy $p(k,s)$1; CAMBI at accuracy $p(k,s)$2; BBAND at accuracy $p(k,s)$3; and VMAF_BA at accuracy $p(k,s)$4 (Chen et al., 2023).

For LIVE-YT-Banding, these numbers support different kinds of expectations. CAMBI’s reported behavior suggests that it is a promising candidate for video-level correlation with subjective banding visibility when the dataset contains 8-bit encodes of originally higher-bit-depth sources and the viewing geometry is properly aligned. FS-BAND’s DBI results suggest strong spatial detection of false contours, but the paper explicitly notes that its global scores would need calibration to LIVE-YT-Banding’s subjective ground truth and would likely benefit from temporal pooling and motion masking. BBAND’s video-level design suggests direct applicability to video datasets, but its evidence base does not include LIVE-YT-Banding. These are therefore extrapolations rather than direct benchmark results.

6. Streaming-system context and encrypted-traffic QoE inference

ReCLive addresses a different but related problem: real-time identification of live video streams and inference of their QoE from encrypted traffic behavior. The system analyzes about 23,000 streams from Twitch and YouTube, covering over 1000 hours of playback, and is deployed in an ISP network serving more than 7,000 subscribers. It extracts request counters every 500 ms and chunk-level telemetry, aggregates YouTube activity by $p(k,s)$5, and classifies live versus on-demand streaming with an LSTM operating on rolling 30 s windows (Madanapalli et al., 2021).

For YouTube, the request-count sequence is

$p(k,s)$6

where each $p(k,s)$7 is the number of upstream request packets observed in a 500 ms interval. The classifier uses an LSTM with hidden and cell vectors of size $p(k,s)$8, followed by an MLP with hidden layers of sizes $p(k,s)$9, $k$0, and $k$1, trained with Adam at learning rate $k$2 and binary cross-entropy loss. Reported live-versus-VoD accuracy on YouTube is $k$3 for 10 s windows, $k$4 for 20 s windows, and $k$5 for 30 s windows. For YouTube Live QoE inference on TCP flows, resolution-bin accuracy is $k$6, and stall detection is reported at $k$7 accuracy with $k$8 recall and a $k$9 false-positive rate (Madanapalli et al., 2021).

The operational significance for LIVE-YT-Banding is indirect but substantial. YouTube Live operates in ultra-low-latency and low-latency modes with small playback buffers, approximately 2–5 s and 8–12 s, respectively. The field deployment reports that YouTube sessions are about 98% VoD and about 2% live, and that live QoE degrades during peak hours, with up to about 15% low-definition and about 7% stall rates in peak windows. The same summary states that the sensitivity of these small buffers to transient congestion is the essence of “banding” in that deployment context. This suggests a systems-level perspective in which LIVE-YT-Banding-style research can be connected not only to pixel-domain artifact detectors but also to traffic-level monitoring that explains when live delivery conditions are likely to produce objectionable visual outcomes (Madanapalli et al., 2021).

7. Limitations, misconceptions, and open directions

A common misconception is that general-purpose fidelity metrics are adequate proxies for banding severity. The CAMBI study directly contradicts this on its controlled Netflix dataset by reporting poor correlation of VMAF and PSNR with banding MOS, despite all sequences having $v^\circ\in\{1,2,4,8,16\}$0 and $v^\circ\in\{1,2,4,8,16\}$1 (Tandon et al., 2021). Another misconception is that all banding detectors are intrinsically video-aware. FS-BAND is explicitly image- and patch-level; its application to LIVE-YT-Banding requires frame-wise use, temporal aggregation, and score calibration. BBAND is video-level, but its temporal model is limited to TI weighting. CAMBI samples every 0.5 s and notes that simple averaging can be suboptimal across many shot changes.

The principal methodological limitations are also clear. CAMBI is luminance-centric, does not explicitly model chromatic banding, and may miss cases in which banding is blended into textured regions because it intentionally suppresses textured and edge-rich areas. FS-BAND does not disclose the patch size or the Mumford–Shah parameters $v^\circ\in\{1,2,4,8,16\}$2, does not model temporal phenomena, and may suffer domain shift under motion, scene cuts, dithering, noise, HDR, or bit-depth changes. BBAND does not incorporate an explicit CSF or JND threshold model, is also luminance-centric, and does not report runtime, memory, or hardware complexity. None of the three papers provides a direct LIVE-YT-Banding benchmark result.

Several open directions follow directly from these constraints. For CAMBI, the literature recommends aligning the scale-to-degrees mapping, window size, and display normalization to the actual viewing geometry of the target dataset, and considering per-shot aggregation when shot changes are frequent. For FS-BAND, the literature recommends temporal pooling, motion masking, and a monotonic mapping to subjective MOS when ported to LIVE-YT-Banding. For BBAND, richer temporal modeling beyond TI weighting is identified as future work. More broadly, the published summaries indicate that LIVE-YT-Banding remains best understood as a benchmark nexus: it is a place where artifact-specific no-reference metrics, patch-level learned detectors, and streaming-system QoE models can be compared, but the available evidence still comes largely from adjacent datasets and task formulations rather than from a unified experimental protocol (Chen et al., 2023, Tu et al., 2020).