Short-video Propagation Influence Rating (SPIR) is a recently introduced task designed to predict the long-term propagation of newly uploaded short-videos. Unlike the traditional Short-video Popularity Prediction (SPP) task that merely focuses on a single interactive indicator, SPIR emphasizes comprehensive influence assessment through multi-dimensional interactive indicators. However, the existing SPIR methods suffer from two limitations: (1) They construct video relationships solely based on semantics, ignoring critical temporal correlations. (2) They initialize video features using only visual information, failing to leverage multimodal attributes for reliable semantic correlation modeling. To overcome these limitations, we introduce a Dual-level Hawkes Process Graph Network (DHPGN) that jointly captures temporal-semantic dynamics for accurate influence analysis. Specifically, we begin by designing a Hawkes Process Graph Neural Network (HPGNN) to capture temporal correlations with semantic attention for modeling the propagation process. Second, we introduce a Heterogeneous Attribution Graph Network (HAGN) to fuse the heterogeneous attributes of short-videos. By leveraging both HPGNN and HAGN, we further develop a Dual-level Graph Propagation (DGP) mechanism, enabling a comprehensive learning of the semantic-temporal correlations between videos. Finally, the learned propagation features of short-videos are input into a neural network for propagation influence prediction. Extensive experiments show that DHPGN achieves substantially better performance than state-of-the-art models on the benchmark dataset.
Unmanned Aerial Vehicles (UAVs) are increasingly employed for dynamic scene monitoring. In this context, cross-modality person re-identification (Re-ID) from UAV platforms-especially under unsupervised conditions-has become a crucial task for applications such as all-weather urban surveillance and emergency response. However, UAV viewpoint variations, especially in pitch angle, cause severe person deformation and appearance changes, leading to large intra-class variance in visual features and, in the absence of manual annotations, results in inaccurate pseudo-label assignment and severe noise accumulation across modalities. Although existing methods attempt to bridge the modality gap via pseudo-label association, they still struggle to overcome the large intra-class variance induced by viewpoint and deformation changes, while remaining highly sensitive to pseudo-label noise. To address these issues, we propose a Dual-granularity Consistent Pseudolabel Optimization (DGCPO) framework. DGCPO integrates complementary cues from both sample-level local features and cluster-level global features to learn viewpoint-invariant and modality-invariant identity representations. This design not only alleviates feature shifts caused by viewpoint and deformation changes but, more importantly, enhances the robustness of pseudo-label assignment through dual-granularity consistency constraints, thereby significantly suppressing noise propagation during model training. Experiments on multiple public crossmodality person Re-ID datasets show that DGCPO significantly outperforms existing methods in unsupervised cross-modality person retrieval accuracy.
Hyperspectral image super-resolution (HSI-SR) plays a pivotal role in remote sensing, environmental monitoring, and various geospatial applications. Despite the remarkable progress of deep learning, HSI-SR still suffers from two fundamental challenges. First, existing methods often lack explicit modeling and effective fusion of spatial–spectral guidance, thereby limiting the exploitation of complementary cross-modal information. Second, the unavailability of high-resolution hyperspectral observations during inference introduces a training–inference discrepancy, which further degrades reconstruction accuracy and robustness. To address these issues, this paper proposes a novel guidance-aware and diffusion-generative transformer framework. Specifically, the framework first captures structured guidance representations in latent space via a spatial-spectral guidance extraction strategy. Built upon this representation, a latent-aware guidance transformer integrates guidance into both the attention pathway and a frequency-enhanced feed-forward network through conditional modulation, enabling collaborative local-global dependency modeling. Furthermore, a diffusion-based guidance modeling mechanism is introduced to learn the distribution of spatial-spectral guidance during training and progressively synthesize high-quality guidance during inference. Extensive experiments on four widely-used benchmark datasets demonstrate that the proposed method consistently achieves superior performance in terms of both quantitative metrics and visual quality, significantly validating its effectiveness and robustness for HSI-SR.
Artificial intelligence (AI)-based face-swapping technology, commonly referred to as “Maskers,” has gained widespread application in film production, television, and digital arts. By enabling the rapid synthesis of realistic facial representations, Maskers play a pivotal role in generating hyper-realistic digital humans. However, due to inherent differences in algorithmic design and implementation, substantial quality disparities exist among Maskers, which can significantly affect users' quality-of-experience of Face Swap Videos (FSVs). To address this issue, we conduct a systematic evaluation of 6 representative Maskers and construct the first Face Swap Video Quality Assessment (FSVQA) dataset, comprising 1,200 FSVs that integrate both AI-generated and real facial content. Subjective experiments confirm significant quality differences among various Maskers, specifically identifying 8 common degenerations and revealing their limitations in synthesizing different types of FSV. Moreover, we propose GRIM, a novel objective quality assessment method for FSVs that comprehensively perceives global quality, identity, facial motion, and deepfake detection features. Experimental results demonstrate that the proposed method outperforms existing approaches, providing a reliable quantitative index and valuable guidance for the advancement of Maskers.
Cross-modality registration between images and 3D point clouds is essential for camera-based localization within pre-built maps, yet it faces the inherent 2D-3D domain gap. While learning-based methods often overfit to training data, and even matching-free retrieval approaches still rely on learned embeddings, we introduce a paradigm of explicit geometric bridging and alignment to address this gap fundamentally. Our framework, DepthI2P, first leverages depth estimation models to produce an explicit geometric proxy, then reformulates registration as a pose retrieval task via direct depth-distribution alignment. This design eliminates the need for learning-based cross-modal matching, yielding a strictly training-free pipeline. To enable robust and scale-invariant comparison, we design composite depth consistency metrics that collectively handle scale ambiguity, density variation, and noise. An efficient coarse-to-fine search strategy is designed to ensure a balance between computational efficiency and registration precision. The approach is theoretically grounded, with analysis demonstrating local convexity guarantees near the true pose. Comprehensive evaluations on KITTI and nuScenes benchmarks show that DepthI2P achieves accuracy competitive with state-of-the-art learning methods. Extensive ablation and correlation studies validate the contribution of each metric component and reveal a positive correlation between depth estimation quality and final registration accuracy. Crucially, the framework exhibits superior generalization across diverse datasets without any fine-tuning. Its plug-and-play nature allows it to function both as a standalone localizer and as a refiner to enhance the output of existing methods.
