Aerial object detection, particularly in Unmanned Aerial Vehicle (UAV) scenarios, presents a persistent challenge due to the prevalence of small targets with low-resolution appearances, ambiguous features, and high missed-detection rates. Although existing real-time detectors have made significant progress, achieving an optimal balance between detection accuracy, robustness, and computational efficiency across diverse aerial environments remains difficult. To address these limitations, we propose SPNC-YOLO, a real-time aerial detection framework built on an architecture-optimized version of YOLOv11. This framework integrates three computationally efficient and complementary innovations designed to resolve specific bottlenecks in small object detection. First, we introduce an improved Space-to-Depth Convolution (SPD-Conv) to mitigate the loss of fine-grained spatial information typically caused by strided convolutions, thereby enhancing feature preservation in deeper layers. Second, a dedicated high-resolution P2 detection layer is incorporated to enlarge the receptive field for distant and minute objects, significantly improving cross-scale feature modeling. Third, we reconstruct the C3K2 block by integrating a Selective Boundary-Aware (SBA) unit and replacing standard bottleneck convolutions with Partial Convolutions (PCConv). Extensive experiments on three benchmark aerial datasets—VisDrone2019, UAVDT, and AI-TOD—demonstrate that SPNC-YOLO achieves superior performance compared to the baseline YOLOv11 and other state-of-the-art real-time detectors.
Deep learning has achieved substantial progress in breast cancer detection from mammograms, with recent state-of-the-art methods leveraging multiple views, such as craniocaudal (CC) and mediolateral oblique (MLO), to improve diagnostic accuracy. Most existing approaches pursue performance gains by increasing architectural complexity and rely exclusively on image data. In contrast, clinical practice heavily depends on expert-derived semantic descriptors defined by the Breast Imaging Reporting and Data System (BI-RADS) lexicon, which guide the interpretation of radiologists. This discrepancy raises a fundamental question regarding the relationship between explicit radiological domain knowledge and learned deep visual representations in breast cancer detection. In this work, we systematically investigate the impact of integrating BI-RADS descriptors into deep learning models for mammography-based breast cancer detection. We evaluate dedicated architectures designed to fuse BI-RADS descriptors with image features and compare them against image-only models and BI-RADS–only tree-based baselines. Our results show that incorporating BI-RADS descriptors consistently improves performance across lesion types and evaluation metrics. Notably, relatively simple models augmented with BI-RADS descriptors outperform substantially more complex image-only architectures. Moreover, BI-RADS integration enhances generalization, enabling a single unified model to effectively classify both mass and calcification lesions. These findings demonstrate that combining structured expert knowledge with deep learning yields more accurate, robust, and clinically aligned breast cancer detection systems.
Vehicle re-identification (ReID) under extreme day-night illumination variations poses substantial challenges for intelligent transportation systems. Nighttime images exhibit reduced contrast, headlight glare, and illumination-induced appearance shifts, all of which degrade cross-domain matching. Existing approaches operate predominantly in spatial domains, where appearance features remain coupled to illumination-dependent statistics. Furthermore, frequency-domain components can encode structural cues that are less sensitive to illumination drift, yet current methods lack explicit quantification of uncertainty to assess feature reliability under challenging conditions. We propose Hierarchical Frequency-Aware Cross-Modal Transformer (HFC-Transformer) for day -night vehicle ReID, which extracts illumination-robust structural representations through frequency-domain analysis and uncertainty-guided fusion. The framework combines modality-specific convolutional feature extraction within a shared Transformer encoding setup with frequency-domain structural disentanglement to perform angular-frequency decomposition. This allows us to filter illumination-sensitive interactions and enforces orthogonal alignment for paired day-night observations. An uncertainty-guided fusion module is then used to aggregate the refined frequency components. Comprehensive experiments on DN-Wild and DN-348 day-night benchmarks validate the effectiveness of HFC-Transformer, achieving 73.6
Automatic detection of Coronary Artery Disease (CAD) based on Cardiac Magnetic Resonance Imaging (CMRI) remains challenging due to the severe scarcity of publicly available CAD-positive scan samples. High acquisition costs, various imaging constraints, safety concerns, etc. are the main reasons for the limited availability of CAD-positive CMRI samples. However, CMRI samples with no sign of CAD are available in somewhat adequate numbers. Such a class imbalance of available training samples causes difficulties in training deep learning architectures for robust detection of the disease. Here, we introduce CAD-GAN+, a pathology-aware image data augmentation framework capable of generating synthetic CMRI samples from randomly sampled Gaussian latent vectors. The two-stage pipeline of the proposed framework includes a GAN with frequency-domain regularization. It is trained with morphological pathology priors to produce CMRI slices consistent with CAD exhibiting anatomically plausible disease characteristics. The classifier (ResNet-34 with Squeeze-and-Excitation attention) of the second stage is trained with an entropy-weighted loss and uses a composite CAD-Consistency Score to identify generated samples that are implausible or diagnostically irrelevant. Only high-fidelity samples validated by the classifier are used to augment the training set. Extensive experiments involving a ResNet-34 based baseline classifier trained only on real samples, as well as the same trained using various augmented training sets, show that the proposed CAD-GAN+ framework substantially improves CAD detection performance, producing improvements of up to 20
Single-image dehazing remains challenging due to the spatially global nature of atmospheric scattering, which degrades contrast and suppresses fine structural details. While recent convolutional and transformer-based methods have improved dehazing performance, many approaches operate purely in the spatial domain and implicitly entangle haze-related degradation with structural information, making it challenging to preserve fine details under moderate and dense haze. We propose a frequency-aware multi-scale convolution–transformer network motivated by the observation that haze predominantly affects low-frequency components, whereas edges and textures reside in high-frequency components. Our model separates haze-dominated low-frequency features and structure-rich high-frequency details using a fixed, separable low-pass filtering module derived from CDF 9/7 analysis taps, with the high-frequency component obtained as a residual. Integrated into a coarse-to-fine U-Net encoder–decoder with hybrid convolution–transformer blocks, deterministic high-frequency injection during decoding preserves details while enabling effective global haze removal. Experiments on natural and remote sensing benchmark datasets (NID, RSID, and SateHaze1k) show that the proposed method achieves the best performance among the compared approaches while using significantly fewer parameters.
Calcium imaging has become a cornerstone technique for the study of the nervous system. It allows researchers to visualize intracellular Ca^2+ dynamics, which serve as a proxy for neuronal activation. One primary limitation of this method is the low signal-to-noise ratio (SNR) of the acquired footage, which may obscure the actual neural signals. In recent years, many denoising techniques have been proposed. Among them, self-supervised approaches have emerged because they do not require clean ground-truth data. However, their main drawback is the inability to capture fine details when faced with highly noisy images. In this paper, we introduce Freq2Clean, a lightweight enhancement module calibrated on synthetic data that operates downstream of an existing denoiser. In the Fourier domain, it fuses the magnitude of the temporally averaged video containing high spatial SNR with the denoiser’s output. Then, it reconstructs an enhanced video via inverse transform. Our results show that Freq2Clean consistently improves SNR and structural similarity index (SSIM) across multiple state-of-the-art denoisers on both synthetic and real datasets, and leads to improved downstream ROI segmentation, while preserving temporal dynamics. The source code repository is publicly available at: https://github.com/MrPio/freq2clean
Audio-visual speech enhancement (AVSE) leverages visual cues to improve speech quality, yet most existing methods treat vision merely as static auxiliary information. Current approaches typically either concatenate visual embeddings directly with audio features or rely on frozen, external synchronization experts, resulting in weak coupling and potential domain mismatch. In this paper, we propose Sync-Flow, an intrinsically synchronization-aware one-step generative model. Unlike prior works, we internalize synchronization by enforcing contrastive alignment between the audio denoising encoder and a visual encoder. This mechanism distills visual phonetic priors directly into the audio latent space without relying on external modules. Furthermore, we introduce an asymmetric injection strategy that restricts explicit visual conditioning to the decoder, effectively preventing shortcut learning while guiding generative refinement. Experiments on VoxCeleb2 and LRS3 show that Sync-Flow achieves competitive performance over representative baselines in PESQ, SI-SDR, and ESTOI, suggesting that intrinsic synchronization regularization is beneficial for robust speech restoration and separation.
Unsupervised object discovery methods using pre-trained self-supervised ViTs can localize multiple objects in images, but cannot assign class labels to these objects, limiting their use in class-aware vision tasks. Prior solutions primarily rely on non-learnable clustering methods, such as k-means or spectral clustering, to group ViT-based object-centric feature representations. In contrast, we hypothesize that a parametric learning-based method for semantic grouping can improve precision by reducing discrepancies between unsupervised partitions and actual visual class boundaries. Therefore, this paper proposes a two-stage learning framework, OMaLa, for unsupervised object mask labeling, enabling object-centric semantic segmentation. In the first stage, it optimizes a clustering objective function to learn k-cluster embeddings by leveraging the shared semantic information among the ViT-based object-centric feature representations. The goal is to distill the initial pseudo-segmentation masks for the unlabeled image dataset. In the second stage, an iterative self-training process refines the pseudo-masks using a segmentation network, enhancing model prediction and semantic consistency. Experiments on the PASCAL VOC dataset demonstrate that OMaLa achieves an mIoU of 49.0
Visual Autoregressive (VAR) models face a severe “Memory Wall” on edge devices due to large model size and substantial KV-cache requirements. In this work, we analyze the Infinity VAR family (2B and 8B) and propose a compression pipeline for deployment on constrained NVIDIA Jetson systems. We diagnose critical bottlenecks: activation outliers reaching 353× the median and channel-skewed cache variance. To address this, we propose a hybrid pipeline combining SVDQuant to structurally decouple weight outliers and asymmetric per-channel KV8 quantization. Our approach reduces the Infinity-8B footprint by 64 → 13.3 GB), fitting it on the mid-range Orin NX with a 4.1 × speedup over Flux.1-dev (W4A4), while achieving superior aesthetic alignment (ImageReward 1.13 vs 0.935). Crucially, we also unlock entry-level feasibility for the Infinity-2B, compressing it from 16.0 GB to 7.71 GB to enable deployment on the Orin Nano. These results establish a new efficiency standard for high-fidelity generative AI at the edge. The code is available at https://github.com/Henvezz95/VAR-Compressor .
Cross-view orientation alignment between co-localized satellite imagery and ground-level photos is a fundamental step towards fine-grained multi-modal spatial reasoning, since the orientation of ground-level photos is not generally known but is necessary for extracting object correspondences across both modalities. Existing datasets for training and evaluating orientation alignment models are typically limited to a single geographical region, raising questions about their generalization to diverse environments across the world. We aim to address this limitation via two complementary contributions: a geographically diverse dataset and a zero-shot method that leverages state-of-the-art vision foundation models. We argue that reliable evaluation of cross-view methods requires geographically representative benchmarks and, to this end, we introduce CVGlobal, a geographically balanced dataset spanning five continents with paired satellite and street-view imagery across urban and rural contexts. Along its side, we present ZeSCO: a training-free, zero-shot method for cross-view orientation estimation built on vision foundation models. ZeSCO makes use of the cross-modality feature matching capabilities inherent to self-supervised foundation models and compares visual tokens from both image modalities while incorporating lightweight priors (such as sky segmentation and multi-layer depth stratification) to estimate the relative orientation between images, all while requiring no task or region-specific training.
Anomaly detection and localization (ADL) in point clouds is a rapidly expanding field of 3D computer vision, owing to its importance in robotic manufacturing and automated quality control. Recent ADL methods extract representations of 3D geometries from a test object and compare them to representations of anomaly-free objects. Despite rapid progress, modern ADL techniques still struggle to meet accuracy expectations in safety-critical fields such as aerospace. The main challenges are learning representations that are highly robust for anomaly detection and developing algorithms that accurately and unambiguously compare these representations. We overcame the first challenge by formulating ADL as a semi-supervised contrastive learning problem and developing a deep representation extractor optimized for anomaly detection. For the second challenge, we compare test representations with anomaly-free representations of multiple reference objects, precisely aligned in a common 3D reference frame. Our method establishes a new state of the art on Real3D-AD and Anomaly-Shapenet datasets, achieving a mean area under the ROC curve of 91.2
Automated diabetic retinopathy (DR) detection from fundus images using Deep Learning (DL) faces challenges from non-standardized evaluations across isolated datasets, obstructing fair comparisons and optimal method selection. This study conducts a systematic comparison of multiple DL architectures across four heterogeneous public datasets (APTOS 2019, IDRiD, DDR, Messidor-2) under identical conditions. We perform two complementary experiments to validate existing work while isolating key design impacts like transfer learning, (1) replicating six literature methods to benchmark robustness and (2) an ablation study on four architectures (VGG16, ResNet50, InceptionV3, AlexNet) comparing from-scratch vs. ImageNet fine-tuning. Results show Hybrid MobileNetV2+SVM achieving consistent accuracy (>0.80 across all datasets), while transfer learning universally outperforms from-scratch training. Experiment Code is available at a Github Repository .
Recent advances in digital pathology have enabled zero-shot segmentation of whole slide images (WSIs), yet deploying the models in practice remains challenging due to their limited generalization capability. Under the zero-shot protocols, task-specific fine tuning is unavailable. In this work, we present a finetuning-free framework that builds WSI level segmentation maps from a frozen pathology vision-language model (VLM) by combining prototype-based compression with graph-based label propagation. We first cluster tile embeddings from a reference set into a compact set of morphological prototypes, and perform inductive label propagation on a class–prototype–tile graph to obtain per tile soft scores. We then introduce a per WSI refinement stage that constructs a tile-to-tile semantic kNN graph and performs transductive propagation to promote context-aware predictions within each slide. This design makes label propagation practical for gigapixel scale medical WSIs while preserving a black box VLM access pattern. Experiments on three distinct histopathology datasets demonstrate the advantages of our method over direct zero-shot tile classification and two prototype-based baselines based on k-means clustering and Gaussian mixture models, respectively.
Decoding fine-grained emotional states from electroencephalography (EEG) signals presents a significant challenge due to their inherent non-stationarity and low signal-to-noise ratio. While deep learning has advanced this field, designing an architecture that effectively balances local feature extraction with long-range temporal modeling remains a difficult trade-off. Current Convolutional Neural Networks (CNNs) are computationally efficient but often constrained by limited receptive fields, whereas Transformer-based models capture global context but typically incur high parameter costs and computational overhead. To address this challenge, we propose LS-Mamba, a lightweight hybrid framework. The architecture incorporates a Large-Small Gating (LSG) module to dynamically modulate features, aiming to separate local variations from global trends via dual-branch convolutions. Subsequently, a bidirectional Mamba encoder is employed to model the continuous temporal evolution of emotional states, capturing long-range dependencies with linear computational complexity. The proposed framework achieves efficient modeling with a compact parameter footprint of only 0.06M. Experiments on the SEED and SEED-IV benchmarks demonstrate that LS-Mamba achieves competitive accuracies of 89.96 https://github.com/xtt-lzt/LS-Mamba .
We present a pattern discovery framework that transforms raw tabular data into actionable intelligence. Our system integrates frequent itemset mining with hierarchical clustering, which uses a composite similarity metric combining both structural and instance-level overlap. Critically, we leverage Large Language Models (LLMs) to summarize the discovered patterns in natural language, perform semantic categorization of patterns and automatically generate multi-level narrative reports designed to facilitate access to discovered patterns for stakeholders at varying technical levels. Evaluated on the MIMIC-III Demo clinical database, our framework’s semantic categorization was validated in a blind study by medical experts, who consistently ranked the LLM-based approach highly despite significant inter-expert variability in their own assessments. We further demonstrate the approach is robust across LLM families and that frontier models produce summaries faithful to the underlying patterns. Interactive similarity search capabilities enable exploratory analysis, transforming pattern discovery from one-way extraction into an interactive dialogue with the data. This work provides a scalable, domain-independent solution that effectively transforms structured data points into interpretable narratives, automating the generation of meaningful insights. Our source code (with Supplementary Material) is available here at https://github.com/radiantlogicinc/cluster_rank_summarize .
Scaling language models to longer contexts is essential for capturing rich dependencies across extended discourse. However, naive context extension imposes significant computational and memory burdens, often resulting in inefficiencies during both training and inference. In this work, we propose CCF, a novel context compression framework designed to enable efficient long-context modeling by learning hierarchical latent representations that preserve global semantics while aggressively reducing input redundancy. CCF integrates segment-wise semantic aggregation with key-value memory encoding, forming compact representations that support accurate reconstruction and long-range understanding. To further enhance scalability, we introduce a training-efficient optimization strategy that couples incremental segment decoding with sparse reservoir sampling, substantially reducing memory overhead without degrading performance. Empirical results on multiple long-context language modeling benchmarks demonstrate that CCF achieves competitive perplexity under high compression ratios, and significantly improves throughput and memory efficiency compared to existing approaches. These findings highlight the potential of structured compression for scalable and effective long-context language modeling.
Surface reconstruction from point clouds is a core challenge in computer vision, especially when the data is partial, noisy, cluttered, and unordered. We introduce BaySurf-SANF, a novel end-to-end probabilistic generative framework that reconstructs 3D surfaces from such degraded inputs. Taking a Bayesian approach in a two-step pipeline, it combines Graph Neural Networks and Transformer encoders to extract local features, and shared latent-space shape priors in the form of a conditional normalizing flow, resulting in posterior samples on the surface representation space. To improve expressiveness, we replace widely used affine layers by Affine Dynamic Layers (ADLs), enabling directional data transformations. Additionally, we simultaneously estimate surface normals to ensure geometric consistency between reconstructed surfaces and their normals. Experimental results demonstrate that BaySurf-SANF achieves state-of-the-art performance across datasets with varying noise levels, outperforming existing explicit reconstruction and generative methods. An ablation study highlights the contribution of each module. These results confirm robustness and flexibility. The source code is available at https://github.com/XMa35/BaySurf-SANF .
Accurate 3D object detection from LiDAR point clouds remains challenging due to irregular sparsity and large-scale variation, while practical systems must run in real time. To this end, we present HFFDet, a object detector with hierarchical feature fusion that improves both effectiveness and efficiency within a pillar-to-BEV pipeline. HFFDet introduces Dual-Path Dynamic Self-Attention (DPDA), which computes set-based attention concurrently along orthogonal grid directions to better capture local structures and to expose hardware-friendly parallelism. To handle scale variation with limited cost, we propose the Multi-scale Pillar Representation (MPR) that hierarchically aggregates pillar features into coarser representations, forming multi-resolution BEV features efficiently. On top of these representations, we build a lightweight BEV 2D Backbone and a complementary fusion mechanism that combines fine-detail and coarse-context features in a coarse-to-fine manner. On the nuScenes benchmark, HFFDet achieves 69.2
Multilingual visual speech recognition (VSR) aims to decode spoken content in multiple languages from lip movements using a single model. However, with the introduction of multilingual data, the model’s performance on one or several languages may degrade, a phenomenon known as the “curse of multilinguality”. This phenomenon indicates the difficulty of achieving performance improvements across all languages. In this work, we propose MuS, a Multilingual Synergy framework for multilingual VSR, which aims to alleviate this issue by promoting the learning of phonetic and linguistic shared representations across languages. At the phonetic level, we adopt a unified phoneme set across different languages to facilitate the learning of shared lip movement patterns corresponding to identical phonemes. To further improve cross-lingual phonetic alignment, we introduce an instance-level alignment loss that minimizes the distance between visual and phoneme representations in the latent space. At the linguistic level, we leverage additional text sequences for phoneme-to-grapheme pre-training, enhancing the model’s mapping capabilities within specific linguistic contexts. To mitigate one-to-many mapping ambiguity, we introduce a distribution-level alignment loss that aligns visual representations with context-conditioned phoneme distributions, thereby promoting more robust context-aware visual representation learning. Experiments on bilingual and multilingual datasets show that MuS consistently improves performance across all languages, demonstrating its effectiveness for multilingual VSR.
Sequential Recommendation (SR) aims to capture evolving user preferences to predict users’ subsequent interactions. Recent studies have developed various SR models and achieve great success. Despite these strides, some studies suggest that user interaction behavior can be broadly categorized as exploitation, reflecting consistent interests, or exploration, indicating interest shifts. This behavioral information is crucial for understanding preference dynamics and improving recommendation performance, yet remains underexplored in existing studies. A key challenge lies in identifying whether an interaction behavior corresponds to exploitation or exploration behavior. In this paper, we propose LDEERec, a model-agnostic, data-centric framework that explicitly mines and leverages exploitation and exploration information for SR. Specifically, we first introduce the zero-shot LLM-based Exploration/Exploitation (E/E) Mining method, which transforms abstract E/E definitions into a predictive proxy task. By leveraging the semantic reasoning capability of Large Language Models (LLMs), our method automatically evaluates interaction behaviors as belonging to exploitation or exploration modes. To incorporate this behavioral information, we further propose a Bi-directional Exploration-Exploitation Learning Strategy for SR model training. This strategy employs a conditionally stochastic scoring mechanism to orchestrate two divergent learning strategies: exploit-first strategy and explore-first strategy, which are used to train SR model separately and resulting in two specialized expert models that excel at capturing different aspects of user behavior. Finally, we propose an Ensemble Prediction mechanism synergistically fuses two models’ diverse knowledge for recommendation. Extensive experiments on three real-world datasets across seven state-of-the-art SR models demonstrate that LDEERec consistently achieves notable performance gains.
