No-reference bitstream-layer point cloud quality assessment (PCQA) can be deployed without full decoding at any multimedia network transmission node to achieve real-time quality monitoring. In this work, we focus on the PCQA dedicated to Octree-RAHT encoding mode. First, to address the issue that existing PCQA databases have a small scale and limited distortion levels, we establish the WPC5.0 database which is the first one dedicated to Octree-RAHT encoding mode with a scale of 400 distorted point clouds including 4 geometry multiplied by 5 attribute distortion levels. Then, we propose the first PCQA model dedicated to Octree-RAHT encoding mode by parsing point cloud bitstreams without full decoding. The model introduces texture bits per point (TBPP) to predict texture complexity (TC) and further derives the texture distortion factor. In addition, the position quantization scale (PQS) is used to estimate the geometry distortion factor, which is then integrated into the model along with the texture distortion factor to obtain the proposed PCQA model named streamPCQ-OR. The proposed model has been compared with other advanced PCQA methods on the WPC5.0, BASICS and M-PCCD databases, and experimental results show that our model has excellent performance while having very low computational complexity, providing a reliable choice for time-critical applications. To facilitate subsequent research, the database and source code will be publicly released at https://github.com/qdushl/Waterloo-Point-Cloud-Database-5.0
In broadcast television and satellite communication systems, compression artifacts and transmission errors introduced by encoding and channel transmission severely degrade the user experience, creating an urgent demand for image quality assessment (IQA) before and after transmission. In this paper, a kernel space distribution measurement method for image quality assessment (KSDM-IQA) designed to quantify distortion perception through the distribution of the kernel space under a decoupled architecture is proposed. The architecture adopts a paradigm that separates feature extraction from similarity perception, reducing the dependence on specific feature extractors and enhancing framework flexibility. A spatial compression mechanism is further introduced to calibrate feature distributions and improve sensitivity to critical spatial information. Moreover, a kernel space distribution consistency strategy is proposed, which involves constructing and centering kernel matrices and leveraging the Hilbert-Schmidt norm and kernel alignment to measure similarity, thereby modelling the nonlinear perceptual similarity of deep features. Finally, the bidirectional and multilevel kernel distribution measurements are fused to assess the overall distortion. Experimental results on multiple public datasets demonstrate that KSDM-IQA is strongly competitive. Notably, our method delivers excellent performance on compression artifacts and channel noise distortion types, demonstrating its potential for extension from conventional IQA to communication transmission quality evaluation. The source code is publicly available at the following website https://github.com/Galaxy-123-coder/KSDM-IQA
Currently, most consumer videos in Standard Dynamic Range (SDR) format undergo media compression. High Dynamic Range (HDR) reconstruction from such videos remains challenging due to the existence of block-based, entangled luminance and texture distortions caused by sequential luminance quantization and block-by-block media encoding. Specifically, block-based luminance banding caused by quantization and block blur introduces fake block contours; meanwhile, block-based texture distortions from cross-block moving objects are inconsistent and cause luminance shifts. Facing these mutually incurred, entangled distortions, most existing methods neglect one or both aspects, as well as their block-based nature, and the remaining multi-stage ones suffer from error propagation. Therefore, we propose a novel Parallel Block-level Enhancement HDR Framework (PBE-HDR) for HDR video reconstruction, conducting independent parallel compression-specific, block-based luminance recovery and texture restoration without cross-distortion entanglement. Without texture involvement, it tackles luminance banding via inter-block global Multi-frame Luminance Equalization (MLE) and intra-block progressive luminance compensation guided by a block prior termed Block Luminance Knowledge (BLK). Meanwhile, it employs implicit distorted-texture enhancement through a multi-scale mutual shuffle-and-attention update mechanism, with dual-frame mutual enhancement for luminance-error neutralization. It also employs pixel shuffling combined with channel-wise attention across multiple scales for implicit inter-/intra-block texture composition and reconstruction, which is infeasible for explicit alignment methods. The final adaptive fusion produces the entanglement-refrained results. Extensive experiments demonstrate that the proposed method outperforms state-of-the-art reconstruction methods.
Volumetric video streaming has recently emerged as a key immersive media technology, enabling content consumers and users to have immersive six degrees of freedom (6DoF) experiences. However, smooth streaming of volumetric content with an enhanced sense of presence is challenging because of the large amount of volumetric data, complex pre-processing, unreliable network transmission, and resource-intensive decoding operations. Current streaming solutions often fail to optimize real-time rendering of the volumetric content and therefore lead to poor streaming experiences, further exacerbated by inefficient bitrate adaptation policies. In this paper, we introduce an MPEG-DASH compatible adaptive volumetric video delivery solution that includes: 1)an interactive web-based volumetric video streaming player that incorporates parallel Draco decoding capabilities and employs the WebGL-based Three.js framework to perform level of detail rendering and playback; 2)an MPEG-DASH-compatible volumetric video delivery framework that uses Draco-encoded frames to formulate DASH chunks with Media Presentation Description (MPD) manifest generation and HAProxy-based traffic management for scalable content delivery; and 3)an enhanced Throughput- and Buffer Occupancy-based Adaptation (TBOA+) algorithm that integrates real-time network conditions and buffer state analysis to optimize the quality selection for immersive human-centric volumetric content. Comprehensive experimental evaluations using four volumetric videos (i.e., Longdress, Redandblack, Loot, and Ricardo10) from the 8i Voxelized Full Bodies (8i VFB v2) and MVUB datasets, along with the scaled 5G Irish dataset, reveal that TBOA+ achieves outstanding performance with a 32.73% higher volumetric bitrate, a 21.15% higher average volumetric quality level, and a 31.52% higher throughput efficiency, all while ensuring more stable playback compared to conventional approaches.
Multi-scale feature compression seeks to eliminate task-insensitive redundancy in multi-scale features while maintaining downstream task performance. However, existing methods struggle to dynamically extract inter-scale correlations, leading to insufficient redundancy elimination. Moreover, they typically fail to accurately reconstruct fine-grained spatial details and to adequately preserve key object semantics, limiting their practical coding performance. To address these issues, a Semantic-consistent Holistic Perception Network (SHPNet)-based multi-scale feature compression is proposed, which not only dynamically captures robust inter-scale correlations for compact feature fusion but also significantly enhances the quality of reconstructed features. Specifically, an inter-scale decorrelation-based adaptive fusion module is first designed to reduce task-irrelevant inter-scale redundancy by dynamically mining global and local correlations across scales. Then, a context-guided detail reconstruction module is developed to recover fine-grained spatial details using diverse contextual cues. Finally, we propose a sensitivity-aware signal-semantic fidelity loss that incorporates sensitivity maps to explicitly guide our SHPNet in preserving task-crucial object semantics. Extensive experiments on both instance segmentation and object detection tasks demonstrate that our proposed method achieves superior coding efficiency and significantly outperforms state-of-the-art multi-scale feature compression methods.
