Learning to rank multimodal content—such as images, videos, and text—according to human preferences is a fundamental challenge in multimedia retrieval. Existing preference optimization methods, originally designed for text-only language models, often fail to handle the inherent noise in multimodal human feedback, suffer from poor calibration of relevance scores, and inefficiently utilize limited annotation budgets. To address these issues, we propose DyCa-GRPO (Dynamic Calibration–enhanced Group-wise Preference Optimization) for multimodal retrieval. Our approach introduces (1) dynamic sampling with margin filtering, which constructs training pairs only when human-rated relevance scores differ significantly—reducing noise and improving sample efficiency; and (2) score-calibrated learning, which trains a multimodal ranker to jointly optimize relative ranking and absolute relevance score prediction. Evaluated on COCO and WebVision with human-derived quality annotations, DyCa-GRPO outperforms DPO, PPO, and GRPO by 4.3–6.8% in nDCG@10, reduces expected calibration error (ECE) by 35%, and maintains robustness under 20% label noise. Our work bridges preference optimization and calibrated multimodal retrieval, offering a practical framework for human-aligned search systems.

Knowledge transfer aims to apply existing knowledge to different tasks or new data, and it has extensive applications in multi-domain and Multi-Task Learning. The key to this task is quickly identifying a fine-grained object for knowledge sharing and efficiently transferring knowledge. Current methods, such as fine-tuning, layer-wise parameter sharing, and task-specific adapters, only offer coarse-grained sharing solutions and struggle to effectively search for shared parameters, thus hindering the performance and efficiency of knowledge transfer. To address these issues, we propose Channel-Wise Parameter Sharing (CWPS), a novel fine-grained parameter-sharing method for knowledge transfer, which is efficient for parameter sharing, comprehensive, and plug-and-play. For the coarse-grained problem, we first achieve fine-grained parameter sharing by refining the granularity of shared parameters from the level of layers to the level of neurons. The knowledge learned from previous tasks can be utilized through the explicit composition of the model neurons. Besides, we promote an effective search strategy to minimize computational costs, simplifying the selection of shared weights. In addition, our CWPS has strong composability and generalization ability, which theoretically can be applied to any network consisting of linear and convolution layers. We introduce several datasets in both Incremental Learning and Multi-Task Learning scenarios. Our method has achieved state-of-the-art precision-to-parameter ratio performance with various backbones, demonstrating its efficiency and versatility.

Vision-based 3D object detection (3DOD) gains lots of attention due to its low cost for deployment compared to Lidar-based tasks, while it suffers from labor-expensive data annotations. At the same time, active learning (AL) has shown great potential in reducing annotation costs in related tasks, which can maximize model performance within very limited labeled data. In this paper, we explore active learning for vision-based 3DOD for the first time. Inspired by the entropy analysis, we involve three concerns to characterize the sample informativeness: sample diversity in input space, feature informativeness in BEV space, and result distribution in prediction space. Based on these concerns, we propose a novel AL framework named HMAD, which utilizes Height Modeling and Adaptive Diversity-based sampling for comprehensive informativeness characterization. In HMAD, we first propose a novel height-guided adversarial module in BEV space, which measures the informativeness of height modeling for 2D-to-3D mapping in an adversarial manner. Furthermore, Budget-aware SpatioTemporal diversity Sampling (BSTS) and Class Balance Sampling (CBS) are proposed to adaptively measure the sample informativeness in input and prediction space, respectively. Finally, the three components are integrated into a two-stage sampling strategy, with which the most informative samples can be selected and annotated for the next iteration. Experiments evidence that HMAD achieves comparable performances by only using 50% annotated training data, and can generalize well on different conditions.

Current methods for scene text detection face four primary challenges: deficiencies of the anchor box approach in capturing text shapes, limited modeling capabilities for arbitrary-shaped text, high computational complexity in method representation, and significant post-processing overhead in detection algorithms. Addressing these challenges, we integrate computational geometry to introduce an efficient, interpretable, and shape-capable arbitrary-shaped text instance representation method, AFRG. This method adapts curve fitting through a selection of adaptive point sets based on geometric medians, leading to the development of a novel anchor-free module suitable for arbitrary-shaped text detection, APG-ATD, applied in scene text detection contexts. Furthermore, we propose an algorithm named MSEAP, which employs multidimensional supervision and evaluation of the adaptive point set across the dimensions of positioning, shape, spatial distribution, and classification to learn higher quality adaptive point sets. Building on the APG-ATD module, we further design an arbitrary-shaped text detection algorithm, MRAPG-ATD, based on multidimensional supervision and evaluation of adaptive points to enhance detection performance.

Consistently stylizing and editing 3D objects from multiple viewpoints is crucial for immersive applications such as virtual reality, augmented reality, and digital entertainment. Nevertheless, existing methods frequently face significant challenges, including inconsistent textures, pronounced drifting artifacts, and compromised geometric integrity when rendered from various perspectives. To effectively address these limitations, we introduce S2-Edit3DV, a novel diffusion-guided framework that reframes multi-view 3D objects editing as a temporally coherent video editing problem. By exploiting the robust single-view generative capabilities of SV3D, our approach reliably propagates initial style edits across different viewpoints, substantially mitigating drifting artifacts prevalent in current video-based editing methods. To further enhance semantic precision and structural preservation, we propose two innovative techniques: Attention-based Differential Style Injection (ADSI) and Adaptive Structural-aware Plug-and-Play (AS-PnP). ADSI utilizes attention-driven semantic embeddings for adaptive and precise style injection, effectively reducing semantic hallucinations. AS-PnP strategically modulates stylized latent features, balancing artistic expression with strict structural coherence. Comprehensive evaluations and ablation studies demonstrate that our proposed framework significantly enhances multi-view consistency, preserves fine-grained geometric details, and ensures accurate semantic alignment, showcasing superior performance and practical value for generating high-quality, creatively stylized, and structurally robust objects.

Low-light image enhancement (LLIE) aims to improve the visibility and quality of images captured under poor illumination. However, existing deep enhancement methods often underemphasize computational efficiency, leading to high energy and memory costs. We propose \textbf{Spike-RetinexFormer}, a novel LLIE architecture that synergistically integrates Retinex theory, spiking neural networks (SNNs) and a Transformer-based design. Leveraging sparse spike-driven computation, the model reduces theoretical compute energy and memory traffic relative to ANN counterparts. Across standard benchmarks, the method matches or surpasses strong ANNs (25.50 dB on LOL-v1; 30.37 dB on SDSD-out) with comparable parameters and lower theoretical energy. Our work pioneers the synergistic integration of SNNs into Transformer architectures for LLIE, establishing a compelling pathway toward powerful, energy-efficient low-level vision on resource-constrained platforms.

Hybrid storage solutions have emerged as potent strategies to alleviate the data storage bottlenecks prevalent in blockchain systems. These solutions harness off-chain Storage Services Providers (SP) in conjunction with Authenticated Data Structures (ADS) to ensure data integrity and accuracy. Despite these advancements, the reliance on centralized SPs raises concerns about query correctness, as the integrity of query results depends on the SPs' trustworthiness. Although ADS can verify the integrity of individual data points, they fall short of preventing SPs from omitting valid results. In this paper, we delineate the fundamental distinctions between data retrieval in blockchains and traditional database systems. Drawing upon these insights, we introduce the BPI framework, which employs a suite of validation models that ascertain the inclusion of all valid content in retrieval outcomes, with low overhead. We further present ''Articulated Search'', a query pattern specifically tailored for blockchain environments, which not only enhances retrieval efficiency but also substantially reduces costs during data user updates. Extensive experimental evaluations demonstrate that the BPI framework achieves outstanding scalability and performance in keyword searches within blockchain environments, surpassing EthMB+ and state-of-the-art search databases commonly used in mainstream hybrid storage blockchains (HSB). Notably, the Articulated Search pattern improves query performance by over three orders of magnitude, highlighting its potential as a transformative approach to blockchain query optimization.

Data synthesis methods have shown promising results in general deepfake detection tasks. This is attributed to the inherent blending process in deepfake creation, which leaves behind distinct synthetic artifacts. However, the existence of content-irrelevant artifacts has not been explicitly explored in the deepfake synthesis. Unveiling content-irrelevant synthetic artifacts helps uncover general deepfake features and enhances the generalization capability of detection models. To capture the content-irrelevant synthetic artifacts, we propose a learning framework incorporating a synthesis process for diverse contents and specially designed learning strategies that encourage using content-irrelevant forgery information across deepfake images. From the data perspective, we disentangle the blending operation from face data and propose a universal synthetic module that generates images from various classes with common synthetic artifacts. From the learning perspective, a domain-adaptive learning head is introduced to filter out forgery-irrelevant features and optimize the decision on deepfake face detection. To efficiently learn the content-irrelevant artifacts for detection with a large sampling space, we propose a batch-wise sample selection strategy that actively mines the hard samples based on their effect on the adaptive decision boundary. Extensive cross-dataset experiments show that our method achieves state-of-the-art performance in general deepfake detection.

Monocular 3D human pose estimation presents a considerable challenge owing to the intrinsic depth ambiguity associated with single-camera observations. Existing methods primarily rely on mean per joint position error (MPJPE) loss to train models for the conversion from 2D to 3D coordinates. However, empirical analysis reveals that models trained solely with point-based supervision may produce biomechanically implausible poses or exhibit significant depth ambiguity, even when achieving low MPJPE. This limitation arises from the fact that point-based loss only considers individual joint locations without accounting for inter-joint relationships. Fortunately, edges of human pose encode critical prior knowledge, including skeleton connectivity and biomechanical distributions. Explicitly modeling edge representations enables the model to overcome the constraints associated with point-only approaches, reducing the uncertainty in the optimization process of the 2D-3D inverse mapping and directly constraining depth ambiguity. Therefore, we propose the Graph-Aware Multi-Representation Aggregation (GAMA-Pose) framework that jointly predicts points and edges, with their fusion serving as the final output. To ensure the accuracy of edge predictions and mitigate depth ambiguity, Anti-Depth-Ambiguity Loss (ADA-Loss) is introduced to supervise the properties of edges and give direct supervision on depth ambiguity. Correspondingly, edge-based metrics are proposed to quantify the error of predicted edges. Experiments conducted on Human3.6M and MPI-INF-3DHP datasets demonstrate that GAMA-Pose effectively addresses the limitations of models relying solely on point constraints, mitigates depth ambiguity, enhances the accuracy of both point and edge predictions, and achieves state-of-the-art (SOTA) performance on both datasets.

In large-scale digital service sharing scenarios, given the large number of participating users, frequent cross-domain service interactions, and high-frequency service transactions, to ensure the trustworthiness of digital services, the architecture of the digital service sharing system usually chooses blockchain as its technological foundation. This not only ensures the security and credible deposit of data, but also achieves the credible traceability of data. However, in the current blockchain-based notarization architecture, there may be potential privacy leakage during data transmission, and users are unable to choose the encryption level of their data for blockchain deposition according to their own needs. Moreover, the underlying databases in current applications using blockchain lack convenient retrieval and categorization functionalities. In this work, we propose a trustworthy blockchain solution based on permission management, which implements hierarchical encryption and enables users to flexibly encrypt data according to their needs. Additionally, through the Double Star storage system, while ensuring the reliability of the system, we have also greatly improved the efficiency of data retrieval and classification. Compared to blockchain platforms like XRP and EOS, our solution achieves superior data retrieval and classification efficiency while implementing layered encryption.

Bayesian Optimization is a powerful technique employed to address black-box optimization problems, finding applications in various domains. Meta-Bayesian Optimization (Meta-BO) is a specific approach designed to improve data efficiency by leveraging information from related tasks. In recent years, there has been notable progress in the field of Meta-BO, particularly in surrogate models and acquisition functions that utilize data from related tasks. However, these advancements have predominantly focused on singular aspects of Bayesian optimization, leaving untapped potential in the integration of these two aspects.We propose a novel approach that enables the surrogate model to effectively integrate the acquisition function for Bayesian optimization tasks. This makes the surrogate model to transcend mere function approximation, effectively addressing the aforementioned problem. Taking inspiration from large language models that receive feedback in actual human dialogue tasks, our approach involves pre-training a neural process surrogate model and subsequently leveraging feedback obtained from real Bayesian optimization scenarios to enhance its Bayesian optimization capability. To achieve this, we have extended the Proximal Policy Optimization to utilize feedback derived from Bayesian optimization, incentivizing the pre-trained surrogate model. Our approach has undergone thorough evaluation across diverse models and various benchmark functions. Remarkably, even with minimal incentives, the models exhibit significant advancements in Bayesian optimization, highlighting the effectiveness and robust generalization ability of our proposed method.

Spiking Neural Network (SNN) is a kind of braininspired and event-driven network, which is becoming a promising energy-efficient alternative to Artificial Neural Networks (ANNs). However, the performance of SNNs by direct training is far from satisfactory. Inspired by the idea of Teacher–Student Learning, in this paper, we study a novel learning method named SuperSNN, which utilizes the ANN model to guide the SNN model learning. SuperSNN leverages knowledge distillation to learn comprehensive supervisory information from pre-trained ANN models, rather than solely from labeled data. Unlike previous work that naively matches SNN and ANN’s features without deeply considering the precision mismatch, we propose an indirect relation-based approach, which defines a pairwise-relational loss function and unifies the value scale of ANN and SNN representation vectors, to alleviate the unexpected precision loss. This allows the knowledge of teacher ANNs can be effectively utilized to train student SNNs. The experimental results on three image datasets demonstrate that no matter whether homogeneous or heterogeneous teacher ANNs are used, our proposed SuperSNN can significantly improve the learning of student SNNs with only two time steps.

The Spiking Neural Networks (SNNs), renowned for their bio-inspired operational mechanism and energy efficiency, mirror the human brain’s neural activity. Yet, SNNs face challenges in balancing energy efficiency with the computational demands of advanced tasks. Our research introduces the RTFormer, a novel architecture that embeds Re-parameterized Temporal Sliding Batch Normalization (TSBN) within the Spiking Transformer framework. This innovation optimizes energy usage during inference while ensuring robust computational performance. The crux of RTFormer lies in its integration of reparameterized convolutions and TSBN, achieving an equilibrium between computational prowess and energy conservation. Our experimental results highlight its effectiveness, with RTFormer achieving notable accuracy on standard datasets like ImageNet (80.54%), CIFAR-10 (96.27%), and CIFAR-100 (81.37%), and excelling in neuromorphic datasets such as CIFAR10-DVS (83.6%) and DVS128 (98.61%). These achievements illustrate RTFormer’s versatility and establish its potential in the realm of energy-efficient neural computing.

Sequential recommendation (SR) methods hold a pivotal position in contemporary recommendation systems due to their capacity to capture a user’s evolving interests based on historical interactions. Nevertheless, exsisting methods predominantly focus on leveraging the local contextual information inherent within user-item interaction sequences. Consequently, they usually struggle to capture the high-dimensional connections between items, impeding their ability to learn high-quality user representations. Although hypergraph effectively learns high-dimensional connections between items by capturing beyond-pairwise relationships, existing hypergraph-based recommendation systems fall short in fully leveraging sequential patterns and tend to overlook the global preference among users.To address these challenges, we introduce a novel framework, Hypergraph Convolution for Sequential Recommendation (HC4SR). In our study, we pivot towards the representation of user preferences at each distinct interaction time-step instead of the whole sequence, a strategy particularly crucial for sequential recommendation systems with extensive interaction histories. This granular user representation facilitates accurate and contextually relevant recommendations. Building upon this framework, we strategically implement a cross-entropy loss function across all items. This framework combined with the loss function can significantly enhance our model’s ability to accurately learn and interpret sequential patterns. To further refine our model’s efficacy, we address global user preference variations by integrating a specialized user preference insertion module within the hypergraph convolution framework. The adoption of the LEPorid parameter initialization method is another strategic move, enhancing the model’s proficiency in comprehending item-user embeddings and giving due attention to long-tail items. Extensive experiments on three real-world datasets demonstrate that the proposed method consistently outperforms state-of-the-art methods.

Spiking Neural Networks (SNNs) have indeed shown remarkable promise in the field of computer vision, emerging as a low-energy alternative to traditional Artificial Neural Networks (ANNs). However, SNNs also face several challenges: i) Existing SNNs are not purely additive and involve a substantial amount of floating-point computations, which contradicts the original design intention of adapting to neuromorphic chips; ii) The incorrect positioning of convolutional and pooling layers relative to spiking layers leads to reduced accuracy; iii) Leaky Integrate-and-Fire (LIF) neurons have limited capability in representing local information, which is disadvantageous for downstream visual tasks like semantic segmentation.

The utilization of virtual try-on has gained popularity in the fashion and e-commerce industries as it enables customers to try on clothing virtually before making online purchases. However, existing virtual try-on techniques encounter difficulties in handling complex poses and distortions, which often result in visible misalignments or defects. To overcome these challenges, we propose DiffusionVTON, a virtual try-on framework that employs denoising diffusion models and an Enhanced Garment Guide decoder. Our approach relies on pose keypoints, target models, and clothing images, reducing additional input requirements and mitigating the effects of potentially inaccurate intermediate predictions. The Enhanced Garment Guide decoder enhances the virtual try-on results by incorporating additional garment information into each layer of the decoder, improving image quality and preserving clothing details. Experimental results on the VITON and MPV datasets showcase that our approach surpasses current methods in terms of image quality and fidelity. This enhancement delivers users with realistic and precise virtual try-on experiences.

Summary With the problems of data sharing and information diddling in the field of education, we construct a highly available education consortium blockchain platform to ensure trusted sharing and privacy protection of education data. We employ erasure codes to process blockchain ledger files and optimize the data storage model according to the characteristics of education data, which can reduce the storage volume effectively. A HotStuff consensus algorithm is designed to access the ordering service of Hyperledger Fabric. A suitable educational blockchain network architecture based on the node complexity of education scenarios is proposed to achieve the high availability of the platform. To manage the education blockchain network, we implement the Fabric deployment based on Kubernetes and achieve the goal of including chaincode into Kubernetes environmental management. To improve the resource utilization of chaincode, we explore the new way of chaincode management by the functional computing service. Finally, on the premise of ensuring a 1/2 fault tolerance rate, the total ledger has decreased by 53.56%. Our platform enhanced the Byzantine fault tolerance while ensuring higher efficiency. Experimental results show that our platform is quite suitable for education scenario with many nodes.

Abstract Live performance is an intuitive way to naturally draft the desired motion in the choreographer's mind. In this paper we present a novel approach to choreographing motions by live performance captured with degree of freedom (3‐DOF) accelerometers. The process begins by placing the accelerometers on the user's limbs according to the pre‐specified positions. The computer then recognizes the performed actions using Hidden Markov Model (HMM), which is pre‐trained by the acceleration data samples automatically generated from a pre‐segmented motion capture database. At last, the captured actions are further synthesized with motion retiming and exaggeration based on the acceleration signals from the accelerometers. This method can intuitively rapid‐prototype the choreographed motions for pre‐production of animation, the avatar control in virtual reality and game‐like scenarios, etc. The experimental results show that it can effectively recognize actions with spatial‐time variance, and is easy‐to‐use especially for a novice with little experience. Copyright © 2009 John Wiley & Sons, Ltd.