Human-Centric Quality Assessment and Visual Attention Modeling for Point Clouds addresses a central challenge in immersive three-dimensional media: how to measure, understand, and predict human perceptual experience when interacting with high-fidelity point-cloud content in virtual and mixed-reality environments. Point clouds are increasingly adopted in extended reality (XR), telepresence, free-viewpoint video, and autonomous systems due to their flexibility in representing complex 3D scenes. However, their irregular structure, large data volume, and the coupled effects of geometric and texture distortions make perceptual quality assessment particularly challenging. Widely used image and video quality assessment metrics are known to correlate poorly with human Mean Opinion Scores (MOS) when directly applied to point clouds, highlighting the need for perceptually grounded quality models tailored to this media type.

This dissertation develops methodologies, datasets, and objective metrics that bridge signal processing techniques with human-centred evaluation in order to improve perceptual alignment in Point Cloud Quality Assessment (PCQA). Beyond metric design, the thesis also provides a human-centred experimental framework and prototype system to support subjective quality assessment and visual saliency analysis in immersive environments, enabling controlled yet valid perceptual studies.

The thesis is organized into three complementary parts. The first part focuses on objective PCQA. Novel quality metrics are proposed to capture the combined influence of geometry and texture on perceived quality. A full-reference metric, PointPCA+, and a no-reference model, M3-Unity, are introduced. These approaches employ modalityaware feature representations that explicitly account for the characteristics of point-cloud geometry and color attributes, together with carefully designed similarity measures or learning-based regression models. Extensive evaluations on public benchmark datasets demonstrate that the proposed metrics achieve improved correlation with human MOS compared to state-of-the-art methods, while also exhibiting robustness across different compression schemes and distortion types. These results confirm the importance of modality-specific modeling and perceptually motivated feature design for PCQA.

The second part of the thesis investigates human visual attention for dynamic pointcloud content in immersive environments. Eye-tracking experiments were conducted in six-degrees-of-freedom virtual reality settings to capture gaze behavior during the viewing of dynamic point clouds. Two datasets were constructed: a task-dependent dataset (QAVA-DPC), in which participants performed explicit visual tasks, and a task-free dataset (TF-DPC), designed to capture natural viewing behavior. The experimental design incorporates systematic preprocessing, stimulus normalization, and error profiling of headmounted display eye trackers to ensure data reliability and reproducibility. These datasets enable a detailed analysis of how task demands, motion, and temporal dynamics influence visual saliency and viewing behavior in immersive 3D scenes.

The third part explores the integration of visual saliency into objective quality assessment. By incorporating both ground-truth and predicted saliency maps into PCQA pipelines, the thesis examines how attention-guided feature weighting and perceptual pooling strategies affect quality prediction performance. Experimental results show that saliency-aware approaches can improve predictive accuracy, although the magnitude of improvement depends on the underlying quality metric and pooling strategy. These findings highlight that visual attention is a valuable perceptual cue, but must be integrated in a principled and task-aware manner. Overall, the results demonstrate that attention-aware models can better prioritize perceptually relevant distortions, which is particularly beneficial for applications such as point-cloud compression, adaptive streaming, and immersive media delivery.

In addition to methodological advances, this dissertation contributes a range of resources that support reproducible research and open science for the research community. The gaze-annotated datasets for dynamic point clouds are made available to support reproducible research and facilitate further investigation of visual attention in immersive media. The thesis also documents detailed experimental protocols inspired by ITU recommendations and adapted to extended-reality settings, contributing to ongoing efforts toward standardization and open science in volumetric media research.

Perceptual quality assessment of Dynamic Point Cloud (DPC) contents plays an important role in various Virtual Reality (VR) applications that involve human beings as the end user. Understanding and modeling perceptual quality assessment is greatly enriched by insights from visual attention. However, incorporating aspects of visual attention in DPC quality models is largely unexplored, as ground-truth visual attention data are scarcely available. Besides, testing methods and procedures for collecting visual attention data are still to be agreed on. This article presents a dataset containing subjective opinion scores and visual attention maps of DPCs, collected in a VR environment using eye-tracking technology. Both the quality score and eye-tracking data were collected during a subjective quality assessment experiment, in which subjects were instructed to watch and rate DPCs at various degradation levels under 6 Degrees of Freedom (DoF) inspection, using a head-mounted display. Qualitative interview analysis was also conducted after the experiment. The dataset consists of 50 DPCs, including 5 reference DPCs, with each reference encoded at 3 distortion levels using 3 different codecs (namely G-PCC, V-PCC, CWI-PCL), amounting to a total of 9 degraded version per reference. Additionally, it incorporates 1,000 gaze trials from 40 participants, yielding a total of 15,000 visual attention maps across all the DPCs. We additionally benchmark objective quality metrics originally designed for static point clouds, evaluating their performance in our dataset using two temporal pooling strategies. Furthermore, we employ the visual attention data that are retrieved during our experiment to evaluate whether the performance of widely used objective quality metrics is improved by considering subjective measurements of visual attention. This dataset establishes a link between quality assessment and visual attention within the context of DPC. Moreover, thematic analysis of the interviews helps uncover user behavior and factors impacting perceptual quality for DPC in 6 DoF. This work deepens our understanding of DPC quality assessment and visual attention, driving progress in the realm of VR experiences and perception.

As large multimodal models (LMMs) advance rapidly across diverse multimodal understanding and generation tasks, the need for systematic and reliable evaluation frameworks becomes increasingly critical. To address this need, this survey provides a structured overview of LMM evaluation, centered around two main axes: multimodal evaluation for understanding and generation. (1) For understanding, a dual-perspective framework is introduced to distinguish benchmarks between general capabilities, which emphasize common tasks, and specialized capabilities, which reflect expert-level competence in domain-specific fields. (2) For generation, evaluation is organized by output modality, including image, video, audio, and 3D content. (3) From a community perspective, this survey further highlights authoritative leaderboards and foundational tools that have been instrumental in establishing a comprehensive evaluation ecosystem for LMMs. By unifying general-specialized understanding and modality-specific generation evaluations, this survey clarifies the current landscape and provides guidance for future research in the LMM evaluation field.

This paper introduces an enhanced Point Cloud Quality Assessment (PCQA) metric, termed PointPCA+, as an extension of PointPCA, with a focus on computational simplicity and feature richness. PointPCA+ refines the original PCA-based descriptors by employing Principal Component Analysis (PCA) solely on geometry data; additionally, the texture descriptors are refined through a direct application of the function on YCbCr values, enhancing the efficiency of computation. The metric combines geometry and texture features, capturing local shape and appearance properties, through a learning-based fusion to generate a total quality score. Prior to fusion, a feature selection module is incorporated to identify the most effective features from a proposed super-set. Experimental results demonstrate the high predictive performance of PointPCA+ against subjective ground truth scores obtained from four publicly available datasets. The metric consistently outperforms state-of-the-art solutions, offering valuable insights into the design of similarity measurements and the effectiveness of handcrafted features across various distortion types.

Volumetric video is a key enabler of immersive extended reality (XR) experiences and is often represented using point clouds for their structural simplicity. However, capturing volumetric content through multi-view acquisition and depth sensing poses many challenges, such as occlusions and depth mismatches. To foster research in this field, we introduce a unique dual-quality point cloud dataset, named UVG-CWI-DQPC, which is designed to support the development of point cloud enhancement, compression, and quality assessment. Our dataset includes 12 dynamic sequences captured simultaneously by: 1) a high-end capture system producing high-fidelity point clouds with extensive processing; and 2) a consumer-grade capture system relying on affordable RGB-D cameras, lightweight processing, and open-source tools. For each sequence, our dataset provides ground-truth point clouds from the high-end capture system and raw RGB-D footage from the consumer-grade capture system, along with calibration data and tools for point cloud generation. This dual-quality setup enables direct comparison and benchmarking of algorithms for densification, occlusion removal, registration, and quality enhancement. Our dataset is publicly available under a permissive license to support reproducible research and standardization work in Moving Picture Experts Group (MPEG) and 3rd Generation Partnership Project (3GPP).

As 3D immersive media continues to gain prominence, Point Cloud Quality Assessment (PCQA) is essential for ensuring high-quality user experiences. This paper introduces ViSam-PCQA, a no-reference PCQA metric guided by visual saliency information across three modalities, which facilitates the performance of the quality prediction. Firstly, we project the 3D point cloud to acquire 2D texture, depth, and normal maps. Secondly, we extract the saliency map based on the texture map and refine it with the corresponding depth map. This refined saliency map is used to weight low-level feature maps to highlight perceptually important areas in the texture channel. Thirdly, high-level features from the texture, normal, and depth maps are then processed by a Transformer to capture global and local point cloud representations across the three modalities. Lastly, saliency along with global and local embeddings, are concatenated and processed through a multi-task decoder to derive the final quality scores. Our experiments on the SJTU, WPC, and BASICS datasets show high Spearman rank order correlation coefficients/Pearson linear correlation coefficients of 0.953/0.962, 0.920/0.920 and 0.887/0.936 respectively, demonstrating superior performance compared to current state-of-the-art methods.

Point clouds (PCs) have attracted researchers and developers due to their ability to provide immersive experiences with six degrees of freedom (6DoF). However, there are still several open issues in understanding the Quality of Experience (QoE) and visual attention of end users while experiencing 6DoF volumetric videos. First, encoding and decoding point clouds require a significant amount of both time and computational resources. Second, QoE prediction models for dynamic point clouds in 6DoF have not yet been developed due to the lack of visual quality databases. Third, visual attention in 6DoF is hardly explored, which impedes research into more sophisticated approaches for adaptive streaming of dynamic point clouds. In this work, we provide an open-source Compressed Point cloud dataset with Eye-tracking and Quality assessment in Mixed Reality (ComPEQ - MR). The dataset comprises four compressed dynamic point clouds processed by Moving Picture Experts Group (MPEG) reference tools (i.e., VPCC and GPCC), each with 12 distortion levels. We also conducted subjective tests to assess the quality of the compressed point clouds with different levels of distortion. The rating scores are attached to ComPEQ - MR so that they can be used to develop QoE prediction models in the context of MR environments. Additionally, eye-tracking data for visual saliency is included in this dataset, which is necessary to predict where people look when watching 3D videos in MR experiences. We collected opinion scores and eye-tracking data from 41 participants, resulting in 2132 responses and 164 visual attention maps in total. The dataset is available at https://ftp.itec.aau.at/datasets/ComPEQ-MR/.

Point clouds represent one of the prevalent formats for 3D content. Distortions introduced at various stages in the point cloud processing pipeline affect the visual quality, altering their geometric composition, texture information, or both. Understanding and quantifying the impact of the distortion domain on visual quality is vital to driving rate optimization and guiding post-processing steps to improve the quality of experience. In this paper, we propose a multi-task guided multi-modality no reference metric (M3-Unity), which utilizes 4 types of modalities across attributes and dimensionalities to represent point clouds. An attention mechanism establishes inter/intra associations among 3D/2D patches, which can complement each other, yielding local and global features, to fit the highly nonlinear property of the human vision system. A multi-task decoder involving distortion type classification selects the best association among 4 modalities, aiding the regression task and enabling the in-depth analysis of the interplay between geometrical and textural distortions. Furthermore, our framework design and attention strategy enable us to measure the impact of individual attributes and their combinations, providing insights into how these associations contribute particularly in relation to distortion type. Extensive experimental results on 4 datasets consistently outperform the state-of-the-art metrics by a large margin.