Recent advances in generative models have highlighted the need for robust detectors capable of distinguishing real images from AI-generated images. While existing methods perform well on known generators, their performance often declines when tested with newly emerging or unseen generative models due to overlapping feature embeddings that hinder accurate cross-generator classification. In this paper, we propose Multimodal Discriminative Representation Learning for Generalizable AI-generated Image Detection (MiraGe), a method designed to learn generator-invariant features. Motivated by theoretical insights on intra-class variation minimization and inter-class separation, MiraGe tightly aligns features within the same class while maximizing separation between classes, enhancing feature discriminability. Moreover, we apply multimodal prompt learning to further refine these principles into CLIP, leveraging text embeddings as semantic anchors for effective discriminative representation learning, thereby improving generalizability. Comprehensive experiments across multiple benchmarks show that MiraGe achieves state-of-the-art performance, maintaining robustness even against unseen generators like Sora.
论文检索
输入标题、作者或关键词，从 1,620 篇学术成果中精准定位
Underwater Instance Segmentation (UIS) tasks are crucial for underwater complex scene detection. Mamba, as an emerging state space model with inherently linear complexity and global receptive fields, is highly suitable for processing image segmentation tasks with long sequence features. However, due to the particularity of underwater scenes, there are many challenges in applying Mamba to UIS. The existing fixed-patch scanning mechanism cannot maintain the internal continuity of scanned instances in the presence of severely underwater color distortion and blurred instance boundaries, and the hidden state of the complex underwater background can also inhibit the understanding of instance objects. In this work, we propose the first Mamba-based underwater instance segmentation model UIS-Mamba, and design two innovative modules, Dynamic Tree Scan (DTS) and Hidden State Weaken (HSW), to migrate Mamba to the underwater task. DTS module maintains the continuity of the internal features of the instance objects by allowing the patches to dynamically offset and scale, thereby guiding the minimum spanning tree and providing dynamic local receptive fields. HSW module suppresses the interference of complex backgrounds and effectively focuses the information flow of state propagation to the instances themselves through the Ncut-based hidden state weakening mechanism. Experimental results show that UIS-Mamba achieves state-of-the-art performance on both UIIS and USIS10K datasets, while maintaining a low number of parameters and computational complexity. Code is available at https://github.com/Maricalce/UIS-Mamba.
The attention mechanism has become a dominant operator in point cloud learning, but its quadratic complexity leads to limited inter-point interactions, hindering long-range dependency modeling between objects. Due to excellent long-range modeling capability with linear complexity, the selective state space model (S6), as the core of Mamba, has been exploited in point cloud learning for long-range dependency interactions over the entire point cloud. Despite some significant progress, related works still suffer from imperfect point cloud serialization and lack of locality learning. To this end, we explore a state space model-based point cloud network termed HydraMamba to address the above challenges. Specifically, we design a shuffle serialization strategy, making unordered point sets better adapted to the causal nature of S6. Meanwhile, to overcome the deficiency of existing techniques in locality learning, we propose a ConvBiS6 layer, which is capable of capturing local geometries and global context dependencies synergistically. Besides, we propose MHS6 by extending the multi-head design to S6, further enhancing its modeling capability. HydraMamba achieves state-of-the-art results on various tasks at both object-level and scene-level. The code is available at https://github.com/Point-Cloud-Learning/HydraMamba.
Image manipulation localization (IML) refers to the task of identifying regions in images that have been altered by specific tampering techniques, such as copy-move, splicing, or inpainting. Transformers have been applied to IML tasks due to their ability to model long-range correlations between pixels through the self-attention mechanism. However, the inherent self-attention mechanism may not accurately model or sufficiently enhance these correlations, particularly in detailed edge traces, due to its limitations. In this paper, we propose an Edge-Aware Affinity Enhancement approach for the IML task. Specifically, we introduce an Affinity Regularization Module to establish inter-patch correlations for feature regularization via random walk propagation. Based on the extracted correlation representation, we propose an Edge-Affinity Guidance strategy to further refine the correlation accuracy, particularly in ambiguous edge regions. Extensive experimental results demonstrate that our method outperforms state-of-the-art image manipulation localization techniques in terms of localization accuracy.
Temporal Intention Localization (TIL) is crucial for video surveillance, focusing on identifying varying levels of suspicious intention to enhance security monitoring. However, existing discrete classification methods fail to capture the continuous progression of suspicious intentions, limiting early intervention and explainability. In this paper, we reconceptualize hidden intention modeling by shifting from discrete classification to continuous regression and propose Suspicion Progression Analysis Network (SPAN), which capture the fluctuations and progression of hidden intentions over time. Specifically, when analyzing the temporal progression of suspicion, we discover that suspicion exhibits long-term dependency and cumulative effects across extended sequences, characteristics significantly similar to the settings in Temporal Point Process (TPP) theory. Based on these insights, we formalize a suspicion score formula that models continuous changes while accounting for temporal characteristics. We also propose Suspicion Coefficient Modulation to adjust suspicion coefficients using multimodal information, reflecting different effects of suspicious actions. Notably, we introduce a Concept-Anchored Mapping method to quantify associations between suspicious actions and predefined intention concepts, enabling understanding of not just actions occurring but also their potential underlying intentions. Extensive experiments on the HAI dataset show that SPAN significantly outperforms existing methods, reducing MSE by 19.8% and improving average mAP by 1.78%,. Notably, SPAN achieves a 2.74% mAP gain in low-frequency cases, indicating superior capability in capturing subtle behavioral changes.Compared to discrete classification systems, out continuous suspicion modeling method enables earlier detection and more proactive interventions, substantially enhancing both system explainability and practical utility in security applications.
Table structure recognition (TSR), the task of extracting logical and physical structures from table images, is critical for document understanding. Current end-to-end image-to-text methods typically employ a top-down strategy where physical structure prediction depends on the logical decoder's output sequence. However, this process often suffers from training instability and misalignment between predicted bounding boxes and ground-truth cell positions. To address this issue, we propose G2LFormer, a novel transformer-based framework that employs a ''Global-to-Local'' query enhancement strategy. Specifically, G2LFormer introduces a Vision-guided Query Enhancer to integrate both textual and visual modalities, significantly improving the overall query representation capability and boosting prediction accuracy. Additionally, we design a Multi-scale Manhattan Vision-guider that leverages a spatial attenuation matrix to guide each query towards its corresponding cell location, effectively balancing local and global information for more precise bounding box generation. Extensive experiments on benchmark datasets demonstrate G2LFormer's superior performance, while ablation studies confirming the significant contribution of each proposed module in achieving state-of-the-art results. The source code and model have been released at: https://github.com/Hzbupahaozi/G2LFormer.
Transparent object segmentation from a single image has been investigated for several years. However, detecting transparent areas from video has not been well explored, especially for different kinds of transparent categories besides glass, due to the scarcity of such a dataset. Therefore, in this paper, we propose the video-based transparent object segmentation task and introduce the first-of-its-kind corresponding dataset named TransVid, which contains nearly 400 videos with a total of 18,523 frames. Based on TranVid, we further propose a new method called TranSeg, in which we innovatively introduce Graph Neural Networks into the temporal segmentation task and combined with a novel Diffusion Model to make the model's segmentation results more accurate. Experimental results show that TranSeg achieves higher accuracy with fewer parameters than previous state-of-the-art models, demonstrating the effectiveness of our method. Moreover, comprehensive ablation analysis reveal several fascinating insights and suggest viable paths for further research.
DS-Det: Single-Query Paradigm and Attention Disentangled Learning for Flexible Object Detection
PDF ↗Popular transformer detectors have achieved promising performance through query-based learning using attention mechanisms. However, the roles of existing decoder query types (e.g., content query and positional query) are still underexplored. These queries are generally predefined with a fixed number (fixed-query), which limits their flexibility. We find that the learning of these fixed-query is impaired by Recurrent Opposing in Teractions (ROT) between two attention operations: Self-Attention (query-to-query) and Cross-Attention (query-to-encoder), thereby degrading decoder efficiency. Furthermore, "query ambiguity" arises when shared-weight decoder layers are processed with both one-to-one and one-to-many label assignments during training, violating DETR's one-to-one matching principle. To address these challenges, we propose DS-Det, a more efficient detector capable of detecting a flexible number of objects in images. Specifically, we reformulate and introduce a new unified Single-Query paradigm for decoder modeling, transforming the fixed-query into flexible. Furthermore, we propose a simplified decoder framework through attention disentangled learning: locating boxes with Cross-Attention (one-to-many process), deduplicating predictions with Self-Attention (one-to-one process), addressing ''query ambiguity'' and ''ROT'' issues directly, and enhancing decoder efficiency. We further introduce a unified PoCoo loss that leverages box size priors to prioritize query learning on hard samples such as small objects. Extensive experiments across five different backbone models on COCO2017 and WiderPerson datasets demonstrate the general effectiveness and superiority of DS-Det. The source codes are available at https://github.com/Med-Process/DS-Det/.
Autonomous vehicles generate massive volumes of point cloud data, yet only a subset is relevant for specific tasks such as collision detection, traffic analysis, or congestion monitoring. Effectively querying this data is essential to enable targeted analytics. In this work, we formalize point cloud querying by defining three core query types: RETRIEVAL, COUNT, and AGGREGATION, each aligned with distinct analytical scenarios. All these queries rely heavily on accurate object counts to produce meaningful results, making precise object counting a critical component of query execution. Prior work has focused on indexing techniques for 2D video data, assuming detection models provide accurate counting information. However, when applied to 3D point cloud data, state-of-the-art detection models often fail to generate reliable object counts, leading to substantial errors in query results. To address this limitation, we propose CounterNet, a heatmap-based network designed for accurate object counting in large-scale point cloud data. Rather than focusing on accurate object localization, CounterNet detects object presence by finding object centers to improve counting accuracy. We further enhance its performance with a feature map partitioning strategy using overlapping regions, enabling better handling of both small and large objects in complex traffic scenes. To adapt to varying frame characteristics, we introduce a per-frame dynamic model selection strategy that selects the most effective configuration for each input. Evaluations on three real-world autonomous vehicle datasets show that CounterNet improves counting accuracy by 5% to 20% across object categories, resulting in more reliable query outcomes across all supported query types.
Scene flow estimation using 4D millimeter-wave radar has emerged as a prominent research focus for 3D dynamic perception. However, compared to LiDAR point clouds, the drastic sparsity of radar point clouds poses challenges in enforcing local rigidity constraints, which are crucial for accurate 3D motion estimation. To address this issue, we propose a novel Gaussian-based pseudo-point generation method that fully leverages two distinct yet complementary data modalities, 3D coordinates and Doppler velocity, to support multi-body rigidity assumptions, effectively capturing fine-grained and structured motion patterns from highly sparse radar point clouds. Furthermore, a velocity calibration mechanism is designed to improve the reliability of fine-grained rigid motion velocity estimation. In addition, a progressive fusion strategy is introduced to systematically integrate fine-grained rigid motion priors at multiple levels, enhancing the robustness of matching costs and motion features while effectively compensating for coarse flows. Experimental results on real-world radar scans from the View-of-Delft (VoD) dataset demonstrate the promising performance of our FGRFlow compared to other leading 4D radar-based approaches, validating the advantages of our design choices.
Radiologists rely on eye movements to navigate and interpret medical images. A trained radiologist possesses knowledge about the potential diseases that may be present in the images and, when searching, follows a mental checklist to locate them using their gaze. This is a key observation, yet existing models fail to capture the underlying intent behind each fixation. In this paper, we introduce a deep learning-based approach, RadGazeIntent, designed to model this behavior: having an intention to find something and actively searching for it. Our transformer-based architecture processes both the temporal and spatial dimensions of gaze data, transforming fine-grained fixation features into coarse, meaningful representations of diagnostic intent to interpret radiologists' goals. To capture the nuances of radiologists' varied intention-driven behaviors, we process existing medical eye-tracking datasets to create three intention-labeled subsets: RadSeq (Systematic Sequential Search), RadExplore (Uncertainty-driven Exploration), and RadHybrid (Hybrid Pattern). Experimental results demonstrate RadGazeIntent's ability to predict which findings radiologists are examining at specific moments, outperforming baseline methods across all intention-labeled datasets.
2D images and 3D point clouds are foundational data types for multimedia applications, including real-time video analysis, augmented reality (AR), and 3D scene understanding. Class-incremental semantic segmentation (CSS) requires incrementally learning new semantic categories while retaining prior knowledge. Existing methods typically rely on computationally expensive training based on stochastic gradient descent, employing complex regularization or exemplar replay. However, stochastic gradient descent-based approaches inevitably update the model's weights for past knowledge, leading to catastrophic forgetting, a problem exacerbated by pixel/point-level granularity. To address these challenges, we propose CFSSeg, a novel exemplar-free approach that leverages a closed-form solution, offering a practical and theoretically grounded solution for continual semantic segmentation tasks. This eliminates the need for iterative gradient-based optimization and storage of past data, requiring only a single pass through new samples per step. It not only enhances computational efficiency but also provides a practical solution for dynamic, privacy-sensitive multimedia environments. Extensive experiments on 2D and 3D benchmark datasets such as Pascal VOC2012, S3DIS, and ScanNet demonstrate CFSSeg's superior performance.
Box-supervised instance segmentation methods aim to achieve instance segmentation with only box annotations. Recent methods have demonstrated the effectiveness of acquiring high-quality pseudo masks under the teacher-student framework. Building upon this foundation, we propose a BoxSeg framework involving two novel and general modules named the Quality-Aware Module (QAM) and the Peer-assisted Copy-paste (PC). The QAM obtains high-quality pseudo masks and better measures the mask quality to help reduce the effect of noisy masks, by leveraging the quality-aware multi-mask complementation mechanism. The PC imitates Peer-Assisted Learning to further improve the quality of the low-quality masks with the guidance of the obtained high-quality pseudo masks. Theoretical and experimental analyses demonstrate the proposed QAM and PC are effective. Extensive experimental results show the superiority of our BoxSeg over the state-of-the-art methods, and illustrate the QAM and PC can be applied to improve other models.
Event cameras with bio-inspired neuromorphic sensors are highly sensitive to brightness changes. When there are moving objects in a scene under constant lighting, event cameras only record motion information and output a sequence of events asynchronously. However, the common flickering light sources, such as fluorescent or LED lamps powered by alternating current exist in various real-world scenarios. When operating under a flickering light source, event cameras output numerous redundant event signals that are triggered by the flickering effect, which overwhelm the useful signals that encode motion information. In this paper, we propose EDeF-Net, an Event streams DeFlickering Network that effectively leverages the spatio-temporal correlation of event streams by modeling both the inter-channel temporal attention and inter-patch spatial attention. To facilitate network training and evaluation, we synthesize the first dataset containing paired flickering and flicker-free event streams. Moreover, we demonstrate that event streams filtered by EDeF-Net yield performance improvements on down-stream applications such as event-based optical flow estimation and object tracking.
This paper introduces OmniGSE, a novel general speech enhancement (GSE) framework designed to mitigate the diverse distortions that speech signals encounter in real-world scenarios. These distortions include background noise, reverberation, bandwidth limitations, signal clipping, and network packet loss. Existing methods typically focus on optimizing for a single type of distortion, often struggling to effectively handle the simultaneous presence of multiple distortions in complex scenarios. OmniGSE bridges this gap by integrating the strengths of discriminative and generative approaches through a two-stage architecture that enables cross-domain collaborative optimization. In the first stage, continuous features are enhanced using a lightweight channel-split NAC-RoFormer. In the second stage, discrete tokens are generated to reconstruct high-quality speech through language models. Specifically, we designed a hierarchical language model structure consisting of a RootLM and multiple BranchLMs. The RootLM models general acoustic features across codebook layers, while the BranchLMs explicitly capture the progressive relationships between different codebook levels. Experimental results demonstrate that OmniGSE surpasses existing models across multiple benchmarks, particularly excelling in scenarios involving compound distortions. These findings underscore the framework's potential for robust and versatile speech enhancement in real-world applications.
Event-based image reconstruction has achieved remarkable progress, benefiting from the high temporal resolution and high dynamic range of event cameras. However, most event-based methods focus on enhancing sRGB image quality, neglecting the potential of leveraging event data for RAW-to-sRGB conversion. Due to the limitations of camera sensors, images processed through standard ISP pipelines often suffer from motion blur and color distortion in dynamic scenes. In contrast, RAW images preserve uncompressed scene information, integrating event signals at this stage enables finer texture recovery and more accurate color correction. To tackle these challenges, we propose EvRAW, a novel event-assisted RAW-to-sRGB image reconstruction network that integrates event signals to promote high-fidelity sRGB image reconstruction. Specifically, we introduce a Motion-guided Structural Enhancement (MSE) module that extracts motion patterns from event streams and aggregates dynamic features to restore fine textures. Additionally, we propose an Adaptive Color Correction (ACC) module that performs region-wise gamma correction and channel-wise color decoding to enhance color fidelity under complex lighting conditions. To evaluate performance in challenging real-world scenarios, we collect a pixel-aligned RAW-Event dataset specifically for this task. Extensive experiments demonstrate that EvRAW achieves state-of-the-art performance in RAW-to-sRGB reconstruction on both synthetic and real-world datasets.
Point cloud regression localization technology has a wide range of applications in the multimedia field. For example, in virtual reality and augmented reality, accurate point cloud localization can significantly enhance the user experience. Recently, point cloud pose regression algorithms based on APR (Absolute Pose Regression) and SCR (Scene Coordinate Regression) have achieved near sub-meter accuracy, requiring multiple repetitive trajectories for training. The key to their success lies in the diversity of viewpoints, temporal changes, and trajectories, which is resource-consuming. However, due to the errors in GPS/INS, the coupling between trajectories is not ideal, and the stability of re-localization is insufficient. Since LiDAR has covered most of the scene, single-shot localization has the potential to approach or even surpass multi-trajectory localization methods through pose enhancement. Specifically, we present Pose Enhancement Localization (PELoc), which feeds one trajectory, proposing SSDA (Single-shot Data Augmentation) and LTI (LiDAR Trajectories-coupled Interpolation) to simulate different driving poses, and we introduce KP-CL (Key Points Contrastive Learning) through feature perturbation to mitigate the differences in viewpoint/temporal phase transformations in similar scenes across different trajectories. Our algorithm has been tested on the Oxford, QE-Oxford, and NCLT datasets, where single-shot localization accuracy can approach near sub-meter level on QE-Oxford and NCLT. The code will be published in https://github.com/Eaton2022/PELoc.
Recent latent diffusion models (LDMs) have been explored to generate diverse domain-specific images based on source domain data, showing promising performance in domain generalization tasks. However, although the generated images present counterfactual augmentation, such as the background and style changes, the distortion of object details disrupts the causal factors, such as texture and shape. This leads to negative outcomes when directly applying LDM to domain generalization in object detection. To address the problems mentioned above, we propose Object-Preserving Counterfactual Diffusion augmentation method (OPCD) to explore the diffusion model to generate diverse domain-specific images without disrupting the object details. First, we construct a region-aware image generation framework, which leverages labeled source domain data to guide LDM in generating region-constrained images that preserve the semantic consistency of the original source images. Second, we propose object-preserving counterfactual augmentation, which retains the object region of the generated image and fuses diversified global information. This ensures that object details are not distorted and that the generated information is maintained. Third, to reduce the resource burden of generating a large number of images in LDM, we design a random insertion strategy. It mixes generated and source domain images, turning limited diversity samples into abundant training data. Experimental results on several benchmark datasets show that OPCD outperforms existing methods in single-domain generalized object detection. Codes can be found at https://github.com/qinhongda8/OPCD.
Multimodal Domain Generalization (MMDG) aims to enhance the robustness of multimodal models against distribution shifts in unseen target domains. Unlike unimodal domain generalization methods, which primarily focus on mitigating domain bias within individual modalities, MMDG faces unique challenges, notably modality heterogeneity (divergent feature spaces) and stability discrepancy (varying sensitivity to domain shifts). To tackle these challenges, we propose Modality-Domain Joint Adversarial Training, a unified framework that addresses these challenges through two key innovations: (1) a tri-discriminator adversarial module that mitigates domain biases in both modality-specific and multimodal representations, while suppressing modality-heterogeneous patterns in the representation space; and (2) a stability-aware dynamic weighting mechanism that adaptively balances modality contributions based on cross-domain stability, reducing reliance on unstable modalities. Additionally, we provide the first theoretical error bound for MMDG, offering a theoretical foundation that supports the effectiveness of our approach. Our approach achieves state-of-the-art performance on the EPIC-Kitchens and HAC datasets while using 75.2% fewer parameters than previous MMDG methods. The source code is available at https://github.com/lihongzhao99/MMDG-Joint-Adversarial-Training.
Live-cell imaging is a powerful tool for studying dynamic subcellular processes by capturing the spatiotemporal organization of the biological microenvironment. However, limitations due to phototoxicity and photobleaching prevent microscopes from achieving high frame rates and high-quality images. Although current deep learning methods can enhance both frame rates and image resolution without compromising cell health, they often overlook the continuity of subcellular trajectories, which leads to discontinuous temporal modeling. It also incurs prohibitive computational costs due to exhaustive correlation computation that hinder real-time applications. To address these issues with high efficiency, we propose Trajectory Space-Time Super-Resolution (T-STSR), a method designed to boost frame rates and resolution in fast subcellular imaging while significantly reducing computational overhead. Our approach incorporates Spatial-Temporal Trajectory Modeling (STTM), which learns a state-space model over spatiotemporal slices to reconstruct particle trajectories at low cost. In addition, our novel Trajectory-Aware Loss randomly subsamples trajectory data during training, promoting continuous trajectory representation and mitigating noise with minimal additional computation. We validated T-STSR on both synthesized and real-world datasets with various particle types and noise conditions, demonstrating that our method achieves superior restoration results while saving 75% inference time compared to the previous SOTA model.