Assessor360: Multi-sequence Network for Blind Omnidirectional Image Quality Assessment

1 Introduction

• Panorama quality metrics matter because degradation can make image content visually unpleasant and undermine high-quality visual content.The motivation is especially relevant as VR viewers experience 360-degree imagery through head-mounted displays.Source: §1 · p. 1
• Conventional viewport methods project ERP images into undeformed viewports, but do not model the sequential browsing process that forms viewport sequences.ODIs commonly use ERP, whose geometric deformation varies by latitude.Source: §1 · pp. 1–2
• Fixed-order samplers produce identical sequences from a given starting point, while existing scanpath predictors target undistorted ODIs and high-level regions.The authors note that quality assessors attend to both low-quality and high-detail regions in distorted ODIs.Source: §1 · p. 2
• RPS combines semantic scene and local distortion information to generate different pseudo-viewport sequences from the same starting point.It uses Equator-guided and Details-guided sampled probabilities.Source: §1 · p. 2
• MFA with DAB fuses viewport semantic and distortion features, while TMM models temporal transitions between viewports.TMM uses a GRU module and MLP layers to model transitions and regress sequence features to a score.Source: §1 · pp. 2–3
• Assessor360 models BOIQA with multiple viewport sequences and was applied to datasets with and without real scanpaths, outperforming state-of-the-art methods.The authors describe it as simulating authentic ODI scoring with multiple scanpaths.Source: §1 · pp. 2–3

2 Related Work

• Traditional 2D IQA metrics adapted to panoramas are poorly aligned with perceived quality, while reference-based methods are hindered by the scarcity of pristine ODI references in authentic scenarios.The passage distinguishes full-reference, reduced-reference, and no-reference OIQA, with the latter also called blind OIQA.Source: §2 · p. 3
• Deep BOIQA methods sample viewports and regress their features, but the surveyed methods do not model sequences arising from multiple observers’ browsing.Examples include cubemap feature aggregation, salient-region tangent viewports, and graph connections between viewports.Source: §2 · p. 3
• Viewport sampling includes uniform unordered, crucial-region unordered, and fixed-direction sequential strategies, but fixed-order sampling differs from evaluators’ variable scanpaths.Uniform coverage may include non-significant regions; crucial-region sampling emphasizes image content but omits sequential order.Source: §2 · p. 3

3.1 Overall Framework

• RPS, MFA, and TMM process an ODI into viewport-sequence quality scores, which are averaged to predict the ODI’s quality.The framework initializes multiple starting points, generates viewport sequences, aggregates features for each viewport, and models temporal transitions.Source: §3.1 · pp. 3–4, Figure 2 · p. 4

3.2 Recursive Probability Sampling

• RPS recursively samples viewport transitions by combining equator-guided probabilities with entropy-based detail probabilities.It uses an equator prior and inhibition of return in ESP, while DSP derives probabilities from patch entropy; their integrated distribution selects subsequent viewports.Source: §3.2 · pp. 4–5

3.3 Multi-scale Feature Aggregation

• MFA fuses shallow distortion-sensitive and deeper semantic features, applying DAB to emphasize local degradation in shallow features.DAB uses a 3×3 convolution followed by channel-attention operations; the fused representation is sent to TMM.Source: §3.3 · p. 5

3.4 Temporal Modeling Module

• TMM uses a GRU to model viewport transitions and regresses its final sequence representation to a quality score.The design reflects the recency effect: recently viewed viewports are described as more likely to affect overall quality judgments, especially during prolonged exploration.Source: §3.4 · p. 6

4.2 Comparing with the State-of-the-art Methods

• Assessor360 improves over state-of-the-art methods across all datasets, including 3% in SRCC over VGCN on OIQA, 1.2% in PLCC on CVIQD, and up to 1.6% in SRCC over AHGCN.The reported gains do not require training with the 2D IQA dataset used by VGCN.Source: §4.2 · pp. 6–7

4.3 Cross-Dataset Evaluation

• Cross-dataset tests report superior generalization; training on CVIQD and testing on MVAQD yields PLCC values 0.5 and 0.3 above MC360IQA and VGCN, respectively.The evaluation also reports approximate SRCC margins of 0.8 over MC360IQA and 0.5 over VGCN when trained on MVAQD and tested on IQA-ODI; CVIQD and MVAQD differ in distortion coverage.Source: §4.3 · p. 7, Table 2 · p. 7, Table 3 · p. 7

4.4 Effectiveness of Recursive Probability Sampling

• RPS-generated viewports outperform ScanGAN360- and ScanDMM-generated sequences for quality prediction, while replacing VGCN sampling with RPS raises MVAQD SRCC by 0.07.The sequence-generation comparison uses OIQA and MVAQD; the VGCN sampling comparison uses IQA-ODI and MVAQD.Source: Table 4 · p. 7, §4.4 · p. 8, Table 7 · p. 9
• RPS improves LEV by 1.75 and 1.82 over ScanDMM and REC by 1.14 and 0.93 over ScanGAN360, while remaining comparable on DTW across JUFE and JXUFE.RPS viewport positions concentrate near the equator and align closely with ground-truth positions, unlike the more dispersed learning-based positions.Source: Table 5 · p. 8, Figure 3 · p. 8, §4.4 · p. 9

4.5 Ablation Studies

• Changing starting-point latitude causes a substantial performance decline, whereas longitudinal shifts cause relatively minor fluctuations; equator-near starts better match typical human browsing behavior.Source: §4.5 · p. 9, Table 6 · p. 9
• MFA and DAB substantially contribute to performance with a minor parameter increase, and the GRU-based TMM represents temporal transitions better than global average pooling.Source: §4.5 · p. 9, Table 8 · p. 9
• As viewport-sequence length increases, performance trends downward for all tested sequence counts, while multiple sequences capture broader scene perspectives and improve robustness.The study tests N = 1, 3, and 5 on MVAQD.Source: §4.5 · p. 10, Figure 4 · p. 10

4.6 Proximate to Human-Observed Performance

• On JUFE, pseudo-sequences generated by RPS contribute nearly as much as ground-truth sequences, though the authors note substantial room to better incorporate human sequences.Source: Table 9 · p. 9, §4.6 · p. 10

• Assessor360 combines RPS, MFA with DAB, and TMM to generate viewport sequences, fuse distorted and semantic features, and learn viewport transitions.Source: §5 · p. 10
• The method demonstrates high performance on multiple OIQA datasets, and RPS effectiveness is validated against two advanced learning-based scanpath-prediction models.Source: §5 · p. 10
• RPS uses fixed transition directions and distances, producing equally spaced viewports.Source: §5 · p. 10

• CVIQD, OIQA, IQA-ODI, and MVAQD lack real scanpath data, whereas JUFE and JXUFE include actual user scanpath coordinates.For datasets with scanpaths, viewport sequences can be sampled directly; RPS generates pseudo sequences when scanpath information is unavailable.Source: A Details of OIQA Datasets · p. 15
• Each JUFE and JXUFE ODI has 300 HMD-recorded viewport coordinates and “good” or “bad” starting points.Source: A Details of OIQA Datasets · p. 15
• Ratings at 5 and 15 seconds under both starting conditions yield four MOS labels per ODI: 5s-bad, 5s-good, 15s-bad, and 15s-good.Source: A Details of OIQA Datasets · p. 15

B Calculation Process of the Scanpath Prediction Metric

• Higher REC and lower LEV/DTW indicate better scanpath prediction; the Human Baseline compares each ground-truth sequence against all ground-truth sequences.The setup sets the number of pseudo sequences equal to the number of ground-truth sequences.Source: B Calculation Process of the Scanpath… · p. 15

C Functions of the Features at Different Stages of Swin Transformer

• The first two Swin Transformer stages primarily capture local distortions, while the latter two emphasize viewport semantics as receptive fields grow.Deep-layer features are used as abstract semantic information.Source: C Functions of the Features at… · p. 15

D.1 Difficulties in Applying Learnable Scanpath Prediction Methods

• Small OIQA datasets, missing scanpath labels, and large learnable-model parameter loads complicate training and can lead to overfitting or hardware-constrained optimization.The learnable scanpath models comprise at least 21% of Assessor360's parameters.Source: §D.1 · pp. 15–16, Table 11 · p. 16

D.2 Advantages of RPS

• RPS generates multiple panorama-based viewport sequences, supports any starting point, and can produce different routes from the same start.The method is presented as simulating multi-assessor scoring and enabling quality assessment from diverse perspectives.Source: §D.2 · p. 16

D.3 Limitation of RPS

• RPS uses fixed transition direction and distance, so generated viewports are equally spaced; future work aims to enable flexible transitions.The authors report that experiments confirmed RPS’s effectiveness for OIQA, while identifying this fixed-spacing design as a limitation.Source: §D.3 · p. 16

D.4 More Observations

• Learning-based NR methods designed for 2D images fail on panorama scenes, while traditional NR methods such as NIQE and BRISQUE perform robustly across datasets.The authors link the 2D methods’ failure to ODI’s high-resolution characteristics.Source: §D.4 · p. 16
• WS-PSNR, WS-SSIM, VIF, MP-BOIQA, and MC360IQA handle homogeneous compression distortions effectively but are weaker on multiple distortion types in OIQA.The OIQA distortions named include JPEG, JP2K, GB, and GN; Assessor360 is reported to perform well in these complex scenarios.Source: §D.4 · pp. 16–17
• Swin Transformer’s first two feature-map stages emphasize local distortions, while its latter two stages emphasize viewport semantics as receptive fields increase.The visualization covers human, landscape, and interior viewport contents.Source: Figure 5 · p. 17

E.1 Effectiveness of DSP

• 17.79 and 40.53 are the DTW score improvements with DSP on JUFE and JXUFE, respectively.The authors motivate DSP by assigning higher sampling probabilities to directions corresponding to texture-rich, high-entropy regions.Source: §E.1 · p. 17
• Entropy maps capture high-detail regions, and entropy-guided sampling directions align with actual observational directions.The paper notes that entropy guides sampling probability; the maximum-entropy value is not itself used as the sampling direction.Source: §E.1 · p. 17, Figure 6 · p. 18

E.2 Time Series Modeling Modules in TMM

• GRU is the best tested TMM alternative on MVAQD and OIQA, outperforming temporal pooling in temporal representation while using fewer parameters than LSTM and Transformer.The authors cite efficient training with limited data and shorter sequences, with reduced overfitting risk, as reasons for selecting GRU.Source: §E.2 · pp. 17–18

E.3 More Visualization

• Assessor360 shows strong fitting ability on CVIQD and IQA-ODI, where existing methods fit IQA-ODI less well; the comparison is shown in Figures 8 and 9.The authors attribute the weaker IQA-ODI fit to its greater content richness and resolution, and say Assessor360 gives more consistent results by modeling scores across perspectives.Source: §E.3 · p. 18, Figure 8 · p. 20, Figure 9 · p. 21
• RPS concentrates viewports near the equator and on distortion regions, unlike ScanGAN360 and ScanDMM, whose sampled regions are more dispersed and can neglect distortion and detail.The paper attributes equatorial focus to ESP and distortion-region focus to DPS; Figure 10 compares generated viewport positions.Source: §E.3 · p. 18, Figure 10 · p. 22
• RPS generates multiple viewport sequences from different starting points and also generates multiple sequences from the same starting point.Figure 7 illustrates both settings.Source: Figure 7 · p. 19