FGLive-10K: Livestreaming BIQA Dataset

• FGLive-10K is a livestreaming BIQA dataset that provides multi-attribute MOS scores and fine-grained pairwise annotations to guide precise camera tuning.
• The dataset comprises over 10K high-resolution images from 555 scenes with systematic variations in 7 camera parameters, supporting both regression and ranking tasks.
• TuningIQA, leveraging human-aware feature extraction and graph-based camera parameter fusion, achieves state-of-the-art performance on fine-grained quality assessment and tuning.

FGLive-10K is a fine-grained blind image quality assessment (BIQA) dataset built specifically for livestreaming camera tuning. It was introduced to address a limitation of conventional BIQA practice: existing datasets and models typically provide only a single coarse overall quality score, whereas livestreaming camera tuning requires supervision that can distinguish subtle but operationally important differences induced by parameter changes such as exposure, ISO, white balance, contrast, saturation, and sharpness. In the formulation accompanying TuningIQA, FGLive-10K couples overall MOS-style labels with fine-grained quality attributes and pairwise preference annotations so that models can learn not only how good an image is, but also why it is good or bad and which of two near-similar images is preferred for tuning decisions (Sheng et al., 25 Aug 2025).

1. Definition and problem setting

FGLive-10K is described as the first dataset of its kind for livestreaming camera tuning (Sheng et al., 25 Aug 2025). Its design is motivated by two properties of livestreaming image quality assessment that differ from conventional IQA settings. First, livestreaming exhibits human-region priority: faces and bodies are usually the most important regions, and degradations affecting them can dominate perceptual quality judgments. Second, livestreaming exhibits parameter interdependency: camera controls are not independent, and exposure-related variables such as ISO, shutter speed, and aperture interact physically rather than additively (Sheng et al., 25 Aug 2025).

Within this problem setting, score-only BIQA is treated as insufficiently granular. Two images may have nearly the same MOS, yet one may be preferable for optimization because of subtle differences in face clarity, exposure balance, or noise. FGLive-10K therefore supports not only regression against absolute quality labels but also fine-grained ranking through pairwise comparison. This suggests a shift from coarse perceptual assessment toward actionable perceptual supervision for camera parameter selection (Sheng et al., 25 Aug 2025).

2. Dataset composition and capture structure

FGLive-10K contains 10,185 high-resolution images at 1920 × 1080 resolution drawn from 555 distinct scenes and captured under systematic camera-parameter variations (Sheng et al., 25 Aug 2025). The associated metadata considers 7 camera parameters: aperture, shutter speed, ISO, white balance, contrast, saturation, and sharpness (Sheng et al., 25 Aug 2025).

A metadata-rich subset, FineLive-p, is defined for explicit camera-parameter interdependence modeling. FineLive-p contains 5,559 training images and 1,148 test images, and includes the full 7-parameter camera metadata (Sheng et al., 25 Aug 2025). In the TuningIQA study, this subset is used specifically to examine the effect of camera-parameter-aware modeling rather than generic image-only BIQA.

The scene-level train/test partition is also explicit. FGLive-10K is split by scene into 451 training scenes with 8,148 images and 72,843 pairs, and 104 test scenes with 2,037 images and 3,705 fine-grained pairs (Sheng et al., 25 Aug 2025). Because the split is scene-based rather than image-based, it is intended to prevent leakage and to test generalization across unseen livestreaming scenarios.

3. Annotation schema and fine-grained supervision

FGLive-10K provides two principal forms of supervision: multi-attribute MOS annotations and pairwise fine-grained preference annotations (Sheng et al., 25 Aug 2025).

For the MOS-style component, annotators assign discrete scores from 1 to 5 for overall quality, face quality, sharpness, exposure, and noise. After post-screening, each image received at least 16 annotations, yielding 50,925 multi-attribute quality annotations in total (Sheng et al., 25 Aug 2025). The attribute semantics are explicitly defined as follows:

Attribute	Meaning
Overall	Global perceptual quality
Face quality	Perceptual quality of human facial regions
Sharpness	Clarity and detail preservation
Exposure	Appropriateness of brightness and contrast
Noise	Absence of grain and visual artifacts

These labels are more informative than a single coarse BIQA score because they identify which perceptual factor is affected. In a camera-tuning context, that distinction is operationally important: the preferred correction for poor exposure is not the same as the preferred correction for poor sharpness or excessive noise.

For subtle comparisons, FGLive-10K supplements MOS with pairwise annotation. The dataset contains 91,946 annotated pairs, of which 19,234 pairs were refined through human pairwise comparison; the paper states that 21% of all pairs were refined in this way (Sheng et al., 25 Aug 2025). The preference construction rule is

$$c_{pq} = \begin{cases} c_{pq}^*, & \text{if } \Delta s > 0.8, \ \frac{1}{K}\sum_{k=1}^K \psi_k(I_p, I_q), & \text{if } \Delta s \leq 0.8, \end{cases}$$

where $c_{pq}^* = \mathbb{I}(s_p > s_q)$ is the initial MOS-based preference, $\Delta s = |s_p - s_q|$, $\psi_k \in \{0, 0.5, 1\}$ encodes worse/equal/better judgment, and $K$ is the number of pair annotations (Sheng et al., 25 Aug 2025). This mechanism explicitly distinguishes coarse MOS-derived preference from human-verified fine-grained preference.

4. Statistical characteristics and motivation evidence

The dataset analysis reported with FGLive-10K provides empirical support for the claim that fine-grained supervision is necessary (Sheng et al., 25 Aug 2025). The overall quality distribution is fairly uniform, indicating coverage across diverse quality levels. Face quality correlates strongly with overall quality, which is consistent with the human-region priority that motivates the dataset. By contrast, sharpness, exposure, and noise are less correlated with each other, suggesting that these factors are complementary and should be modeled separately (Sheng et al., 25 Aug 2025).

A central observation concerns annotation reliability at different granularity levels. Annotator correlation with MOS is high on average, with PLCC = 0.85, but within narrow MOS tiers it declines sharply; in the 4.2–5.0 tier it drops to PLCC = 0.24 (Sheng et al., 25 Aug 2025). This is used to justify pairwise refinement: when images are already of relatively high quality and differ only subtly, absolute scalar MOS becomes a weak supervisory signal for preference-sensitive optimization.

Taken together, these observations support two conclusions. First, human-centric perceptual structure matters disproportionately in livestreaming. Second, relative judgments are better suited than single-image MOS for small quality differences. A plausible implication is that FGLive-10K is not merely a larger BIQA corpus, but a dataset intended to alter the learning target itself—from coarse score prediction to decision-oriented quality modeling.

5. Role in TuningIQA

FGLive-10K is the empirical foundation for TuningIQA, a BIQA framework designed for livestreaming camera tuning with both single-image mode and pairwise mode (Sheng et al., 25 Aug 2025). TuningIQA integrates Human-aware Feature Extraction (HFE) and Graph-based Camera Parameter Fusion (GCPF), and is trained with a multi-task objective combining regression and ranking.

HFE is designed to emphasize the most perceptually important regions, especially human subjects. Its pipeline uses Faster R-CNN to detect humans, a Human Aesthetics Network built on EfficientNetV2-M to extract multi-scale features, ROI Align to extract human-region features, nine spatial partitions centered around the human subject, and a cross-attention fusion between human-aware features and backbone features (Sheng et al., 25 Aug 2025). The partition update is expressed as

$$\mathbf{F}_k = \Phi_k(\hat{\mathbf{M}}_k) + \hat{\mathbf{M}}_k,\quad k \in \{1,\ldots,9\},$$

and the cross-attention layer as

$$\mathbf{F}_q = \text{CrossAttention}(\mathbf{Q}, \mathbf{K}, \mathbf{V}) = \text{softmax}\left(\frac{\mathbf{Q}\mathbf{K}^T}{\sqrt{d_k}}\right)\mathbf{V},$$

with $\mathbf{Q} = \mathbf{F}_h \mathbf{W}_Q$, $\mathbf{K} = \mathbf{F}_b \mathbf{W}_K$, and $\mathbf{V} = \mathbf{F}_b \mathbf{W}_V$ (Sheng et al., 25 Aug 2025).

GCPF is used when camera metadata is available, especially on FineLive-p. It defines a heterogeneous graph $c_{pq}^* = \mathbb{I}(s_p > s_q)$0 with 8 nodes: one visual node and seven parameter nodes corresponding to the camera settings (Sheng et al., 25 Aug 2025). The graph includes cross-modal links, Exposure triangle connections among ISO, shutter speed, and aperture, a Post-processing chain among contrast, saturation, and sharpness, and a Color correlation relation in which white balance affects saturation perception (Sheng et al., 25 Aug 2025). Reasoning is performed with two GAT layers,

$c_{pq}^* = \mathbb{I}(s_p > s_q)$1

with the final fused feature taken from the updated visual node (Sheng et al., 25 Aug 2025).

The learning objective combines confidence-weighted regression and fine-grained ranking:

$c_{pq}^* = \mathbb{I}(s_p > s_q)$2

with

$c_{pq}^* = \mathbb{I}(s_p > s_q)$3

so that ranking is emphasized more strongly than regression (Sheng et al., 25 Aug 2025). This weighting is consistent with the dataset’s stated purpose: camera tuning depends heavily on choosing between near-similar alternatives.

6. Empirical results and practical tuning performance

On FGLive-10K, TuningIQA is reported to outperform state-of-the-art BIQA baselines in both score regression and fine-grained quality ranking (Sheng et al., 25 Aug 2025). The main comparison includes the following results:

Method	SRCC	PLCC	FG-ACC
TuningIQA (MobileNetV3-S)	0.9308	0.9302	0.7065
TuningIQA (ResNet-50)	0.9385	0.9364	0.7284
LIQE	0.9235	0.9206	0.6533
MUSIQ	0.8662	0.8687	0.6861
MT-A	0.8821	0.8829	0.6223
CLIP-IQA+	0.7056	0.7039	0.5570

The reported interpretation is that some baselines achieve reasonable regression but still perform poorly on fine-grained ranking, indicating that score prediction alone is not sufficient for camera tuning (Sheng et al., 25 Aug 2025).

The attribute-wise results further support this point. TuningIQA leads across the reported dimensions with Sharpness at SRCC 0.935, PLCC 0.963, FG-ACC 0.773; Noise at SRCC 0.805, PLCC 0.905, FG-ACC 0.746; Exposure at SRCC 0.897, PLCC 0.949, FG-ACC 0.776; and Face at SRCC 0.885, PLCC 0.880, FG-ACC 0.725 (Sheng et al., 25 Aug 2025). These values are used to argue that the human-aware design is especially effective for face-centric quality understanding.

Ablation evidence is also explicit. On FGLive-10K, adding HFE improves SRCC: 0.8757 → 0.8917, PLCC: 0.9201 → 0.9267, and FG-ACC: 0.7258 → 0.7496. On FGLive-p, adding GCPF improves Baseline SRCC: 0.8206, with GCPF: 0.8552, and with GCPF + HFE: 0.8613, with analogous gains in PLCC and FG-ACC (Sheng et al., 25 Aug 2025). This establishes a functional distinction between the two modules: HFE benefits the full dataset, whereas GCPF is useful when explicit parameter metadata is available.

Practical deployment relevance is assessed by simulating incremental parameter adjustments and using IQA metrics to select the best image. In a subjective comparison across 44 scenarios with 12 volunteers, TuningIQA achieves a 76% win rate vs. LIQE and a 74% win rate vs. MUSIQ (Sheng et al., 25 Aug 2025). This suggests that the dataset’s fine-grained annotation scheme supports not only offline benchmark gains but also better parameter-selection behavior in a tuning loop.

7. Scope, significance, and naming ambiguity

FGLive-10K is practically significant because it provides a dataset tailored to the requirements of livestreaming camera tuning: it captures real camera-parameter variations, includes human-centered perceptual attributes, adds pairwise fine-grained preference labels, supports both regression and ranking learning, and enables research on camera-parameter-aware quality modeling (Sheng et al., 25 Aug 2025). Within BIQA, it therefore occupies a specific niche at the intersection of perceptual modeling, camera control, and livestreaming QoE.

The resource also sharpens a broader methodological point. Traditional BIQA datasets usually provide one MOS score per image; FGLive-10K adds multi-attribute labels and pairwise preferences. This suggests a movement from static quality scoring toward supervision that can guide intervention. In the context of livestreaming, where tuning decisions are incremental and quality differences can be subtle, such supervision is structurally better aligned with the downstream task.

There is, however, a naming ambiguity in later arXiv usage. In "FIGMA: Towards FIne-Grained Music retrievAl" (Anand et al., 4 Jun 2026), the query-side term “FGLive-10K” is described as referring to the FGMCaps 10K test split, a 10,000-pair evaluation set for fine-grained music retrieval rather than a livestreaming BIQA dataset. That paper states that this 10K split is built from MTG-Jamendo, Music4All, and JamendoMaxCaps, and is annotated with Tempo, Key, Chord progression, Beat count, Genre, and Mood for bidirectional text-audio retrieval evaluation (Anand et al., 4 Jun 2026). This is a different resource and task setting. Accordingly, in current technical usage, the primary and unambiguous meaning of FGLive-10K in computer vision and IQA is the livestreaming camera-tuning dataset introduced with TuningIQA, while the music-retrieval usage reflects a separate query-context label rather than the same benchmark (Sheng et al., 25 Aug 2025, Anand et al., 4 Jun 2026).