Multi-user video streaming succeeds only when it delivers both high overall quality of experience (QoE) and fairness. This challenge becomes far more severe in immersive virtual reality (VR) streaming scenarios. Unlike traditional 2D video, equal bandwidth does not imply equal experience: users' diverse interaction patterns lead to sharply different viewport prediction accuracy and thus unequal QoE. Worse still, VR users are markedly sensitive to quality degradation and rebuffering, because even brief negative events can shatter immersion and trigger session abandonment. To tackle these challenges, we propose Mavs, a bandwidth allocation framework for multi-user VR video streaming. At its core, Mavs couples VRweight, a lightweight mechanism that quantifies each user's rate-contribution or reception priority from VR-specific signals, with SafeRL, a safety-aware decision engine that determines a group-level rate-transfer budget under an explicit abandonment constraint. Together, they improve the system's social welfare by 19.7% while relatively reducing abandonment by 65.1%. A real-world multi-user VR streaming prototype further confirms a 17.9% gain in mean opinion score (MOS)-based social welfare.
Reversible data hiding (RDH) in JPEG images has been extensively investigated for grayscale scenarios, while effective RDH for color JPEG images remains largely underdeveloped due to the heterogeneous statistical and perceptual characteristics of the $YC_{b}C_{r}$ compression pipeline. Existing methods typically extend grayscale JPEG-RDH schemes independently to different channels, resulting in limited exploitation of cross-channel redundancy and suboptimal embedding performance. To address this issue, this paper proposes ATME-CJ, an adaptive ternary matrix embedding framework for color JPEG images that enables coordinated reversible embedding across the entire $YC_{b}C_{r}$ domain. The proposed framework jointly exploits intra- and inter-channel dependencies of zero-valued AC coefficients to improve embedding suitability under severely constrained redundancy. Building upon this, a ternary matrix embedding mechanism with perceptual distortion optimization is developed to enhance embedding efficiency while preserving visual fidelity. Furthermore, a channel-aware payload allocation strategy is introduced to adaptively regulate embedding modifications according to channel sensitivity, leading to a more favorable capacity–distortion trade-off. Extensive experiments on multiple benchmark datasets demonstrate that ATME-CJ consistently outperforms state-of-the-art JPEG-RDH extensions for color images in terms of embedding capacity, perceptual quality, and rate-distortion performance, validating the effectiveness of the proposed unified $YC_{b}C_{r}$-aware embedding framework.
Text-video retrieval (TVR) demands precise alignment between dynamic visual content and textual semantics. However, existing methods face a fundamental trade-off where sparse sampling often misses critical inter-frame dependencies, while dense sampling incurs prohibitive computational costs. To resolve this contradiction, we propose a novel $\bf {B^{3}}$ framework, Bridging temporal gap on Brown Bridge via stochastic process based modeling, that models video representation evolution as a Brownian Bridge Process. Unlike conventional discrete point modeling, $\bf {B^{3}}$ encodes latent temporal uncertainty within bounded intervals, enabling the reconstruction of fine-grained continuous video representations directly from sparsely sampled inputs. This design effectively infers missing intermediate information while strictly preserving boundary consistency, restoring realistic and coherent temporal dynamics that are largely neglected in traditional pipelines. To further strengthen cross-modal alignment, we introduce a Text-Guided Temporal Proxy Generation module, which uses textual cues to locate and highlight semantically salient temporal segments in the continuous feature stream, guiding the model to focus on task-relevant visual dynamics. Extensive experiments are conducted on three standard TVR benchmarks, MSRVTT, DiDeMo, and ActivityNet Captions, together with downstream VideoQA datasets including NExT-QA and MSRVTT-QA. The results show that $\bf {B^{3}}$ achieves state-of-the-art retrieval performance with 80% fewer frames than competing methods, while significantly reducing GPU memory consumption and accelerating inference speed. Our framework offers a new and efficient paradigm for balancing temporal representation quality and computational efficiency in cross-modal video understanding tasks.
Synthetic aperture radar (SAR) automatic target recognition (ATR) is a core task in multimedia remote sensing and military reconnaissance owing to the inherent all-weather imaging capabilities of SAR. Under battlefield constraints, noncooperative targets (NTs) are observed only from partial aspect angles (PAA), causing severe out-of-distribution (OoD) mismatch between training and test data. Current data-driven methods struggle to learn robust pose-invariant representations, resulting in poor generalization to unseen viewpoints. To address this challenge, we propose a two-stage cross-modal framework termed Generative Mesh-Image Pre-training (GMIP), which leverages 3D target meshes as a global geometric prior to regularize PAA SAR ATR for improved OoD generalization. Rather than mining available information solely within the SAR image domain, GMIP uses mesh data as a medium for cross-target semantic interaction to learn disentangled representations and transfer their shape components to PAA SAR ATR tasks. In the generative pre-training stage, a cross-modal encoder-decoder learns to explicitly disentangle pose-invariant shape factors and imaging-geometry-related pose factors through interaction-guided mesh-to-image transformation. In the knowledge distillation stage, a bidirectional contrastive knowledge distillation strategy injects mesh-derived shape priors into a generic SAR ATR model to enhance OoD generalization. Extensive experiments on the MSTAR and ATRNet-STAR datasets demonstrate that GMIP outperforms state-of-the-art methods across most PAA settings, including substantial improvements on unseen viewpoints under several evaluation conditions.
Semantic Multiple Object Tracking (SMOT) is an emerging task in the multimedia understanding community, aiming not only to localize objects over time but also to understand their behaviors through instance captioning, video captioning, and interaction recognition. To address the limitations of existing baselines, we propose the Instruction-based Semantic Multi-Object Tracker (ISMOT), which consists of a tracker and a vision-language model fine-tuned with a three-stage supervised strategy. Specifically, we design a unified framework in which task-specific instructions guide the model to perform different semantic understanding tasks. We further adopt a progressive supervised fine-tuning strategy that enables the model to handle multi-task learning and multi-image inputs. Finally, we show that the hidden states learned for semantic understanding can be adapted as person re-identification tokens to enhance tracking accuracy. Experimental results on the public BenSMOT benchmark demonstrate that our compact and mutually reinforcing paradigm outperforms the baseline by a clear margin. Ablation studies and further discussion validate the effectiveness of ISMOT.
