Camouflaged object detection (COD) aims to segment camouflaged objects that are highly similar to their surrounding backgrounds, thereby posing challenges to reliable feature discrimination and accurate contour localization. The emergence of the Segment Anything Model (SAM), which is a visual foundation model pre-trained on large-scale datasets and exhibits strong visual representation and relationship modeling capabilities, offers the potential to address this issue. However, due to the absence of a COD-specific visual corpus in the training data, its direct application to camouflaged object segmentation is susceptible to interference from backgrounds and non-camouflaged objects, leading to limited segmentation performance. To address these limitations, we proposed a dual-path frequency-domain adaptation framework, named Frequency-Enhanced auto-Prompt Adaptation SAM (FEPAS), which synergistically integrates frequency representation embedding and frequency-guided localization prompting to adapt SAM for COD. Specifically, Frequency-Enhanced Adapters (FEAs) are designed to extract and enhance latent frequency-domain cues of SAM features to rectify its inherent limitations regarding frequency-aware perception for better distinguishing camouflaged objects from backgrounds. In addition, a trainable Wavelet Prompt Module (WPM) is built to automatically generate location prompts of camouflaged objects via multi-level learnable wavelet transforms and a dense cross-level fusion strategy, enabling FEPAS to focus on the accurate segmentation of camouflaged regions. Extensive experiments across three widely used COD benchmark datasets verify the effectiveness of the proposed components and demonstrate that FEPAS achieves superior COD performance against other state-of-the-art methods. Furthermore, studies on downstream applications show the promising applicability of FEPAS.
Data incompleteness and privacy preservation are critical challenges in federated incomplete multi-view clustering. Although recent studies have proposed several solutions, they often suffer from widespread view-missing issues and the difficulty of establishing cross-client consensus under privacy constraints, which hinders reliable missing-data recovery and high-quality representation learning. To address these issues, we propose a structure-guided missing-data recovery framework for federated incomplete multi-view scenarios. Specifically, each client constructs a dynamic anchor graph to capture local structural relationships and performs reliable missing-view data recovery via anchor-based reconstruction. Global structural information is introduced to constrain this recovery process, thereby improving the consistency and reliability of the recovered data. Concurrently, a compact-statistics communication mechanism is designed to exchange pseudo-labels and centroid matrices for global consensus building. The resulting global consensus labels guide representation learning and structural alignment, while consensus centroids enhance semantic consistency across clients. Furthermore, the framework avoids explicit construction of dense affinity matrices and achieves linear complexity with respect to the number of samples via implicit computation, significantly reducing communication overhead and computational cost. Extensive experiments on eight benchmark datasets demonstrate that the proposed method achieves superior clustering performance over existing federated methods, and outperforms centralized multi-view clustering approaches in terms of both communication and computational efficiency.
Textured meshes have become an essential representation for immersive multimedia applications, where reliable objective quality assessment is required to guide compression, transmission, and rendering systems. Existing model-based and projection-based methods typically rely on either intrinsic mesh attributes or rendered observations alone. However, model-based approaches may overlook perceptual distortions in the projection domain, while projection-based approaches may fail to capture intrinsic geometric degradations. In this paper, we propose the Cross-Modal Spatial Alignment (CMSA) method, a hybrid structure for full-reference textured mesh quality assessment. CMSA first samples anchor faces from the reference mesh and searches for spatially corresponding local regions on the distorted mesh. For each aligned region, it jointly extracts face-level geometric attributes and multi-modal projection patches, including RGB, depth, and normal information. A geometry encoder and a projection encoder are then employed to encode the intrinsic 3D structure and rendered appearance, respectively. The resulting features are integrated through local gated fusion, refined by reference-guided cross-attention, and finally regressed to the perceptual quality score. Extensive experiments on the TSMD, SJTU-TMQA, and YANA datasets demonstrate that CMSA achieves state-of-the-art or highly competitive performance against model-based, point-based, projection-based, and hybrid quality assessment methods, validating the effectiveness of spatially aligned cross-modal representation for textured mesh quality assessment.
In the rapidly evolving landscape of multimedia system, the ability to seamlessly understand, generate and manipulate video content remains a fragmented challenge, often addressed by disparate, task-specific models. This fragmentation limits the generalization and synergistic capabilities crucial for advanced AI. To overcome this, we propose the MaestroV, a novel holistic model designed to concurrently achieve video understanding, text-to-video generation, and comprehensive video editing functionalities. Technically, MaestroV employs three transformer-based architectures: a Multimodal Large Language Model (MLLM) to interpret user-supplied visual and textual inputs; a Diffusion Transformer (DiT) to produce high-quality video outputs; and a Connector Transformer that acts as a mediator between the MLLM and DiT, transmitting the gathered information from the former to the latter. Upon receiving user inputs, the input video and text are directed to the MLLM model. If the objective is video understanding, the MLLM output is transformed into text using a de-tokenizer. For generation and editing tasks, the learned sequential queries are further processed through the MLLM, guiding it to generate the required information. This information is then transferred to the visual domain by the Connector Transformer. The transferred features are utilized by the DiT as conditions for generating the targeted video. For video editing, we additionally incorporate the VAE-encoded input video into the DiT to obtain granular guidance. A meticulously designed four-stage progressive training approach, coupled with a well-curated multimodal generation dataset, endows MaestroV with unified capabilities. Across various benchmarks, extensive experiments showcase superior performance in both video understanding and generation, along with its robust and precise control over video editing tasks. The source code and model weights will be released.
Deep hashing enables efficient image retrieval by representing images as compact binary codes, but the learned hash space is vulnerable to small adversarial perturbations. Existing untargeted attacks often use a hash-space center to guide the perturbed query toward regions where irrelevant database samples become more likely to be retrieved. Such center-based guidance is effective, yet it mainly provides a query-level attack direction and does not explicitly exploit the category-level semantic structure already encoded in database hash codes. As a result, the adversarial code may still preserve affinity to the original semantic categories. To address this limitation, this paper proposes Semantic Prototype Diversion Attack (SPDA). SPDA first derives a binary semantic prototype for each category by summarizing the hash codes generated by the attacked model. For each query, it then constructs a diversion center that favors semantically dissimilar database samples over similar ones. During adversarial optimization, SPDA attracts the perturbed query code toward the diversion center while repelling it from the prototypes of the query's original labels. This joint supervision pushes the adversarial code toward irrelevant retrieval regions and away from the source-category semantics. Experiments on MIRFlickr, NUS-WIDE, and COCO demonstrate that SPDA achieves stronger untargeted retrieval degradation than representative deep hashing attacks across different code lengths, backbones, hashing models, and adversarially trained settings.
