Offline MARL necessitates policies capable of modelling high-dimensional, multi-modal joint actions from fixed datasets.Unimodal Gaussian policies struggle to represent multiple coordinated action modes, whereas diffusion- and flow-matching approaches, though more expressive, typically require multi-step sampling or a separate teacher-distillation phase for one-step execution.To address these limitations, we introduce SOMP, a one-step generative policy framework for CTDE-compatible offline MARL.SOMP reformulates MeanFlow as an action-space predictor by integrating a history-conditioned Action-to-Action source and agent-level coupling to capture coordination with critic-free per-agent action outputs. In addition, it constrains both ends of the one-step trajectory through inference-endpoint consistency and near-data velocity alignment, ensuring that the direct-prediction objective is well posed. Moreover, a distributional-shift control stack stabilises critic-guided improvement via behavioural anchoring, critic scale regulation, and dataset-statistics-informed actor routing.Evaluated on two benchmark families across 16 datasets, SOMP reconciles policy expressivity with inference efficiency, achieving superior aggregate performance over comparable offline MARL and flow-based baselines while retaining native one-step execution.

Components of the same category within facade images usually have similar semantic features and regular distribution patterns. Exploiting such structural priors can improve parsing model performance in information-deficient scenarios. We propose a pluggable heuristic strategy that enhances existing deep learning frameworks without architectural modifications, requiring only model-agnostic data processing and parameter-free feature map transformations. Our method introduces: (1) a deterministic pre-processing step using a simple segmentation model to evaluate the semantic clarity of input data, integrated with curriculum learning for stable parameter updates; and (2) latent-space feature transformations generating joint response maps to capture structural patterns. Experiments demonstrate that the method can significantly enhance the performance of host models, particularly when processing facade images with incomplete patterns caused by occlusions, low-light conditions, or similar challenges.

The deployment of lightweight object detectors on remote sensing edge platforms is severely constrained by the rigid trade-off between perception capacity and metabolic expenditure. To solve this fundamental challenge, we draw inspiration from the superior energy efficiency of the mammalian brain and the principles of connectomics to introduce CONERSLite. By emulating the dual mode synergy of biological neural systems, CONERSLite integrates a Compact Anatomical Backbone (CAB) representing the stable anatomical connectome and a Functional Connectome Router (FCR) that mimics the plasticity of the functional connectome. Our framework achieves a peak mAP of 82.35% on the DOTA-v1.0 dataset with only 28.3 M parameters and 195 G FLOPs, effectively establishing a new accuracy–efficiency Pareto frontier for remote sensing. On the HRSC2016 dataset, it reaches a state-of-the-art mAP of 98.62% while reducing the total parameter count by approximately 45% compared to high-precision optimized models like RTMDet. These results demonstrate that the application of connectomics principles provides a biologically grounded and highly efficient solution for resource-constrained remote sensing object detection.

We study language-guided target localization in synthetic aperture radar (SAR) imagery for deployment on resource-constrained platforms. Existing vision-language models either rely on heavy backbones unsuitable for edge devices or are designed for natural images, overlooking SAR-specific characteristics such as speckle noise, weak scattering responses, and geometric distortions. The proposed model, TVLightFormer, combines a lightweight dual-modal encoder (MobileNetV3 and TinyBERT) with a grouped-query attention (GQA) mechanism for efficient cross-modal interaction and an activation-free lightweight feature pyramid network (LFPN) to handle scale variation while preserving weak scattering signals. The individual modules are not claimed as newly invented components; the main contribution lies in their SAR-aware integration for edge-oriented cross-modal localization. We evaluate the model on five remote sensing datasets—SOMA-1M, ATRNet-STAR, GAIA, MLRSNet, and SODAS—under a unified localization setting, and we explicitly discuss the limitations introduced by weak or scene-level annotations. The results show that TVLightFormer achieves a favorable trade-off between accuracy and efficiency, reaching an average mIoU of 69.8% with 27.4 M parameters and 9.7 GFLOPs. Ablation studies quantify the contribution of each component. The model is suited for edge-oriented scenarios where computational resources are limited. We also provide a critical analysis of failure cases, SAR-specific disturbance factors, loss-function choices, and dataset-protocol sensitivity.

The well-aligned attribute of CLIP-based models enables its effective application like CLIPscore as a widely adopted image quality assessment metric. However, such a CLIP-based metric is vulnerable for its delicate multimodal alignment. In this work, we propose FoCLIP, a feature-space misalignment framework for fooling CLIP-based image quality metric. Based on the stochastic gradient descent technique, FoCLIP integrates three key components to construct fooling examples: feature alignment as the core module to reduce image-text modality gaps, the score distribution balance module and pixel-guard regularization, which collectively optimize multimodal output equilibrium between CLIPscore performance and image quality. Such a design can be engineered to maximize the CLIPscore predictions across diverse input prompts, despite exhibiting either visual unrecognizability or semantic incongruence with the corresponding adversarial prompts from human perceptual perspectives. Experiments on ten artistic masterpiece prompts and ImageNet subsets demonstrate that optimized images can achieve significant improvement in CLIPscore while preserving high visual fidelity. In addition, we found that grayscale conversion induces significant feature degradation in fooling images, exhibiting noticeable CLIPscore reduction while preserving statistical consistency with original images. Inspired by this phenomenon, we propose a color channel sensitivity-driven tampering detection mechanism that achieves 91

Event cameras capture sparse brightness changes with high temporal resolution and high dynamic range, compensating for the deficiencies of the conventional RGB frames. However, previous multi-modal fusion techniques typically fail to handle the inherent heterogeneity between RGB frames and event streams, thus easily leading to noise amplification or redundant feature integration during cross-modal fusion. In this paper, we propose a Cross-Modal information inTeraction transFormer, coined as CMTFormer, which hierarchically integrates RGB and event information to achieve efficient and stable multimodal collaboration. Specifically, we design a shallow-to-deep information interaction scheme. In the shallow stage, we present the Shallow Alignment Module (SAM) to achieve an efficient fusion of RGB and event low-level features, which mitigates attribute disparities and prevents noisy information. In the middle stage, we devise the Cross-modal Enhancement Module (CEM) that utilizes texture and edge information to produce mutually reinforced middle-level features. In the deep stage, we present the Learnable Deep Fusion Module (LDFM) which performs high-level information aggregation through learnable weights, thus enabling the network to adaptively fuse RGB and event clues. A Spatial Prior Module is further designed to utilize global spatial information to enhance localization accuracy. Extensive experiments are conducted on two prevalent event-based object detection benchmarks, i.e., DSEC-Detection and PKU-DAVIS-SOD. Our CMTFormer consistently surpasses the detection counterparts in both uni-modal and multi-modal settings, strongly demonstrating the effectiveness of our paradigm. Codes will be available upon publication.

Visual place recognition (VPR) aiming at predicting the location of an image based solely on its visual features is a fundamental task in robotics and autonomous systems. Domain variation remains one of the main challenges in VPR and is relatively unexplored. Existing VPR models attempt to achieve domain agnosticism either by training on large-scale datasets that inherently contain some domain variations, or by being specifically adapted to particular target domains. In practice, the former lacks explicit domain supervision, while the latter generalizes poorly to unseen domain shifts. This paper proposes a novel query-based domain-agnostic VPR model called QdaVPR. First, a dual-level adversarial learning framework is designed to encourage domain invariance for both the query features forming the global descriptor and the image features from which these query features are derived. Then, a triplet supervision based on query combinations is designed to enhance the discriminative power of the global descriptors. To support the learning process, we augment a large-scale VPR dataset using style transfer methods, generating various synthetic domains with corresponding domain labels as auxiliary supervision. Extensive experiments show that QdaVPR achieves state-of-the-art performance on multiple VPR benchmarks with significant domain variations. Specifically, it attains the best Recall@1 and Recall@10 on nearly all test scenarios: 93.5

Visual Place Recognition (VPR) is a major challenge for robotics and autonomous systems, with the goal of pre dicting the location of an image based solely on its visual features. State-of-the-art (SOTA) models extract global descriptors using the powerful foundation model DINOv2 as a backbone. These models either explore the cross-image correlation or propose a time-consuming two-stage re-ranking strategy to achieve better performance. However, existing works only utilize the final output of DINOv2, and the current cross-image correlation causes unstable retrieval results. To produce both discriminative and stable global descriptors, this paper proposes a stable cross-image correlation enhanced model for VPR called SciceVPR. This model explores the full poten tial of DINOv2 in providing useful feature representations that implicitly encode valuable contextual knowledge. Specifically, SciceVPR first uses a multi-layer feature fusion module to capture increasingly detailed task-relevant channel and spatial information from the multi-layer output of DINOv2. The designed module not only en hances local feature quality but also reduces feature extraction time by approximately 50 % compared to existing adaptations. Secondly, SciceVPR considers the invariant correlation between images within a batch as valuable knowledge to be distilled into the proposed invariant feature projector. After knowledge distillation, SciceVPR is able to produce a self-enhanced global descriptor from a single input that achieves performance comparable to a multi-frame correlated global descriptor. These two innovations enable SciceVPR to produce fairly robust global features regardless of domain shifts (e.g., changes in illumination, weather and viewpoint between pictures taken in the same place). Experimental results demonstrate that the base variant, SciceVPR-B, achieves SOTA performance on several challenging urban benchmarks. The large variant, SciceVPR-L, performs on par with SOTA two-stage models, scoring over 3 % higher in Recall@1 compared to existing models on the challenging Tokyo24/7 dataset. Our code is available at https://github.com/shuimushan/SciceVPR.

Using images captured by UAVs for high-fidelity 3D building reconstruction in architectural engineering is popular and effective nowadays; however, planning a flight trajectory that maximizes reconstruction quality with minimal flight time remains a critical challenge. This paper proposes a universal co-optimization framework that bridges reconstruction objectives with flight dynamics through an integrated planning paradigm. The proposed approach performs initial flight planning by solving a Traveling Salesman Problem over candidate viewpoints and updating them according to the unit-length contribution criterion. The adaptive radius is determined, and subsequently, the sphere-based corridor is constructed to enforce the trajectory passing all updated viewpoints within the corresponding spatial tolerances. Next, an optimal control problem is formulated and solved using a nonlinear solver to obtain the final flight trajectory satisfying both dynamic and safety constraints. Experimental comparisons with state-of-the-art methods on three public scenes and two real scenes captured by ourselves demonstrate that the proposed approach significantly improves flight efficiency, reducing travel distance and flight duration by approximately 10% to 40% with comparable or superior reconstruction quality.

Weakly-supervised point cloud semantic segmentation (WS-PCS) has attracted increasing attention due to the challenge of sparse annotations. A central problem is how to effectively extract informative features from the annotated points, enabling reliable supervision. Although many existing works extend 2D graph convolution to 3D point cloud data, 2D convolution inherently assumes feature localization, which is an assumption that does not hold in point clouds, and lacks consistent semantic offsets. To address this, we propose a novel Bilateral Graph Convolutional (BGC) method, which refines graph edges into two categories: regular edges and offset edges, providing improved guidance for WS-PCS. Firstly, we create the Local Bilateral Relations (LBR) module to learn the relational features of edges in local point cloud graphs, encompassing both regular and offset edges. To the best of our knowledge, we are the first to utilize offset edges to capture irregular semantic offsets in point cloud data. Secondly, we propose the Adaptive Pooling (AP) module, which adaptively pools edge information learned from LBR, enhancing the feature characterization ability by incorporating salient and pervasive features. Finally, we design BGC as BGC-Net and evaluate its performance against recent networks on four datasets, achieving state-of-the-art results.

Reliable road condition detection using drone imagery is critically important, particularly under harsh weather conditions such as rain, fog, and snow, which cause reduced visibility and blurred objects. Traditional detection methods are limited in effectively handling these severe scenarios due to their static feature extraction approaches. To address these challenges, we propose an innovative affine Lie group convolution and weather-adaptive feature enhancement network (ALGC-WFEMNet). The core innovation of this method lies in the affine Lie group convolution (ALGC), which leverages the mathematical framework of affine Lie groups to introduce a dynamic convolution mechanism. This mechanism adaptively modifies convolution kernels based on affine transformations, significantly enhancing the model's robustness against weather-induced variations in scale, rotation, and visibility. Furthermore, the ALGC framework integrates a learning-based weather condition coefficient, dynamically adjusting kernel responses to specific environmental conditions such as rain, fog, and snow. This theoretical advancement not only emphasizes the mathematical novelty of applying affine Lie groups in convolutional neural networks (CNNs) but also substantially improves feature extraction and adaptability for object detection tasks. Experimental validation on unmanned aerial vehicle (UAV)-based road inspection datasets demonstrates that our ALGC-WFEMNet achieves a mean average precision (mAP) of 60.48%. Furthermore, we deploy the model within a UAV-internet of things (IoT) system to verify its practical effectiveness, achieving an inference time of 23.31 s on a Raspberry Pi.

Large Language Models (LLMs) remain highly vulnerable to diverse attacks, particularly in black-box settings where the internals of target models are inaccessible. Existing black-box defenses typically rely on pre-defined filtering heuristics, which often fail to generalize to unseen attack types and target model architectures. We introduce EvoDefense, an experience-guided co-evolving black-box defense paradigm. EvoDefense employs a guard LLM to detect malicious queries and an experience memory module to accumulate defense knowledge from previous interactions. At the core of EvoDefense is a continuous attack-defense evolution loop, where an attack generator and the guard model iteratively refine their attack strategies and defense policies through experience-guided optimization. This design enables EvoDefense to generalize across unseen attacks and target models without retraining. Experiments on HarmBench, AdvBench, and AlpacaEval show that EvoDefense achieves consistently strong defense performance across seven popular models and five representative LLM attacks, while preserving competitive general capabilities. On HarmBench, EvoDefense reduces the attack success rate (ASR) of AutoDAN-turbo on Gemini-3-flash and LLaMA-3-8B-Instruct from 29.4

Accurate road detection from UAV imagery under adverse weather remains a significant challenge due to reduced visibility, motion blur, and environmental interference. To address these issues, we propose RSFC-EAFANet, a robust detection framework that integrates Robust Scale Fusion Convolution (RSFC) with an Edge-aware Adaptive Feature Aggregation (EAFA) module. RSFC dynamically adjusts convolution kernels to better handle scale variation and degraded visual inputs, while EAFA enhances edge information and suppresses noise in challenging environments. To support evaluation under realistic conditions, we construct a dedicated UAV-based road inspection dataset comprising 19,832 images collected across diverse weather scenarios such as rain, snow, and fog. This dataset features fine-grained annotations and provides a valuable benchmark for assessing detection performance in adverse environments. Extensive experiments on this dataset show that RSFC-EAFANet achieves a mean Average Precision (mAP) of 58.48%, outperforming strong baselines such as NAS-FPN and MegDet by a clear margin. The model also maintains a competitive inference speed of 23.2 FPS, demonstrating a strong balance between accuracy and efficiency. These results highlight both the effectiveness of RSFC-EAFANet and the importance of weather-aware benchmarks for advancing UAV-based road detection. It has made significant contributions to applications in real-world scenarios such as intelligent traffic monitoring, road safety early warning, and road inspection under extreme weather conditions.

Hyperspectral images (HSIs) provide a wealth of information surpassing human visual capabilities, enabling precise identification of remote sensing targets. However, it faces significant challenges, including insufficient long-range dependency modeling, difficulties in data collection, and the tendency of models to get trapped in local optima during training. To overcome these obstacles, we present the sorted texture-aware glance and gaze network (ST-GGNet) tailored for HSI classification. First, we propose the glance and gaze attention (GGA) mechanism, which employs feature interaction-based long-term modeling to minimize information loss across spectral bands and focus on critical land cover features within HSI. Subsequently, the sorted texture-aware module (STM) is introduced to deeply mine and efficiently utilizes detailed texture and spectral information, thereby enhancing accuracy even with limited training data. Additionally, we propose the budding growth optimization algorithm, budding growth optimizer (BGO), which integrates a budding growth mechanism to help the model discover better solutions, boosting optimization and classification performance. Experimental evaluations conducted on four public HSI datasets-Pavia University, Salinas, Houston, and WHU-Longkou-demonstrate the superior performance of ST-GGNet compared to nine state-of-the-art (SOTA) classification methods. Specifically, under limited training samples, ST-GGNet achieves overall accuracies (OAs) of 99.42%, 96.88%, 96.86%, and 97.74%; average accuracies (AAs) of 98.90%, 98.01%, 97.07%, and 92.48%; and Kappa coefficients of 99.24%, 96.53%, 96.59%, and 97.03%, respectively. The findings reveal that ST-GGNet not only maintains strong robustness and generalization but also effectively suppresses noise and excels at distinguishing spatially similar adjacent land covers, especially in low-sample scenarios, consistently outperforming existing SOTA methods. We have released our code and models at https://github.com/Pluviophile-sy/ST-GGNet

State-of-the-art image captioning often relies on supervised models trained on domain-specific image-text pairs, which can be expensive and time-consuming to annotate. Additionally, these models struggle with performance drops when applied to data from different distributions. To overcome these challenges, zero- shot image captioning has emerged, utilizing models trained on unlabeled image and text data. Pre-trained vision-language and generative language models have shown promise in this area, but existing methods do not fully leverage the cross-modal representation capabilities of large-scale models and tend to overlook smaller objects in images. We propose EntroCap, a novel zero-shot image captioning method that incorporates an Entropy-based Retrieval Strategy. During training, our Hierarchical Projector optimizes CLIP's joint embeddings to capture comprehensive global signals with multi-level representations. During inference, the Entropy-based Retrieval Strategy adjusts prediction logits to better include smaller, less frequent objects. Additionally, our Balancing Gate mechanism guides the language model to produce more detailed and contextually relevant captions by balancing local and global signals. Extensive experiments on MSCOCO, Flickr30k, and NoCaps demonstrate that EntroCap outperforms previous zero-shot approaches in both in-domain and cross-domain settings. We also conduct experiments to quantitatively validate EntroCap's effectiveness in describing small objects.

Unmanned Aerial Vehicle (UAV) image semantic segmentation faces challenges in recognizing novel categories due to closed-set training paradigms and the high cost of annotation. While open-vocabulary semantic segmentation (OVSS) leverages vision-language models like CLIP to enable flexible class recognition, existing methods are limited to low-resolution images, hindering their applicability to high-resolution UAV data. Current adaptations—downsampling, cropping, or modifying CLIP—compromise either detail preservation, global context, or computational efficiency. To address these limitations, we propose HR-Seg, the first high-resolution OVSS framework for UAV imagery, which effectively integrates global context from downsampled images with local details from cropped sub-images through a novel cost-volume architecture. We introduce a detail-enhanced encoder with multi-scale embedding and a detail-aware decoder for progressive mask refinement, specifically designed to handle objects of varying sizes in aerial imagery. We evaluated existing OVSS methods alongside HR-Seg, training on the VDD dataset and testing across three benchmarks: VDD, UDD, and UAVid. HR-Seg achieved superior performance with mIoU scores of 89.38, 73.67, and 55.23, respectively, outperforming all compared state-of-the-art OVSS approaches. These results demonstrate HR-Seg’s exceptional capability in processing high-resolution UAV imagery.

Referring Remote Sensing Image Segmentation (RRSIS) aims to accurately locate and segment target objects in high-resolution aerial imagery based on natural language descriptions. Most existing approaches either directly modify Referring Image Segmentation (RIS) frameworks originally designed for natural images or employ image-based foundation models such as SAM to improve segmentation accuracy. However, current RRSIS models still face substantial challenges due to the domain gap between remote sensing and natural images, including large-scale variations, arbitrary object rotations, and complex spatial–linguistic relationships. Consequently, such transfers often lead to weak cross-modal interaction, inaccurate semantic alignment, and reduced localization precision, particularly for small or rotated objects. In addition, approaches that rely on multi-stage alignment pipelines, redundant high-level feature fusion, or the incorporation of large foundation models generally incur substantial computational overhead and training inefficiency, especially when dealing with complex referring expressions in high-resolution remote sensing imagery. To address these challenges, we propose CD2FSAN, a CLIP-driven dynamic feature selection and alignment network that establishes a unified framework for fine-grained cross-modal understanding in remote sensing imagery. This network first follows the principle of maximizing cross-modal information to dynamically select the visual representations most semantically aligned with the language from CLIP’s hierarchical features, thereby strengthening cross-modal correspondence under image domain shifts. It then performs adaptive multi-scale aggregation and alignment to integrate linguistic cues into spatially diverse visual contexts, enabling precise feature fusion across varying object scales. Finally, a dynamic rotation correction decoder with differentiable affine transformation was designed to refine segmentation by compensating for orientation diversity and geometric distortions. Extensive experiments verify that CD2FSAN consistently outperforms existing methods in segmentation accuracy, validating the effectiveness of its core components while maintaining competitive computational efficiency. These results demonstrate the framework’s strong capability to bridge the cross-modal gap between language and remote sensing imagery, highlighting its potential for advancing semantic understanding in vision–language remote sensing tasks.

In open and dynamic environments, object detection is affected by rain, fog, snow, and complex lighting conditions, leading to decreased accuracy and posing a threat to driving safety. Infrared images can provide clear images at nighttime or in adverse weather conditions. Combined with the mature development of existing cross-modality object detection technologies, both of them offer support for addressing object detection issues in adverse weather scenarios. This paper establishes a novel dataset named Adverse Weather and Illumination Dataset (AWID) to simulate intricate real-world scenarios and proposes a cross-modal object detection algorithm for adverse weather scenarios in autonomous driving, named CME-YOLO, which is based on RGB and infrared images. It integrates the Cross-Perception Transformer Fusion algorithm, CPTFusion, and the Adaptive upsampling technique, AdSample, to enhance the extraction of detailed information and supplement effective information. CPTFusion fuses features from different modalities through multi-scale feature extraction and optimal fusion strategy computation. AdSample adaptively improves the utilization of key features and the quality of the resulting feature tensor. Experiments on two public datasets and AWID show that CME-YOLO performs optimally, with an mAP50 value on the FLIR dataset 6.8% higher than the state-of-the-art MPFT algorithm, verifying its excellent performance in autonomous driving object detection tasks.

Using images captured by Unmanned Aerial Vehicles (UAVs) to perform 3D reconstruction is a cost-effective way to acquire high-quality 3D models for large-scale urban scenes. The challenge, however, has become choosing camera viewpoints and planning flight path accordingly. Existing methods either plan the aerial path heuristically or train a reconstructability predictor, where the accuracy and completeness losses are optimized separately in a two-phase approach, leading to inaccurate reconstruction results and poor generalizability. To address these issues, this paper proposes a dual-task learning framework that establishes the correlation between viewpoint poses and reconstruction quality. In particular, the reconstructability estimation problem is modeled as two subtasks: reconstruction accuracy and reconstruction completeness, allowing both subtasks to be tackled simultaneously within a unified network. The model’s generalizability is improved by the soft parameter sharing and a new dual-loss function with trainable weight parameters. In addition, an adaptive viewpoint optimization strategy is proposed to refine an initial set of viewpoints generated based on the learned reconstructability. Our framework is extensively evaluated on both public datasets and two datasets we collected. Qualitative and quantitative experimental results demonstrate the superiority of our method in both synthetic and real scenes. Our framework achieves consistent improvements over state-of-the-art approaches by an average of 3.6% in F-score with 15% fewer images, and surpasses Oblique Photography with a 6% F-score gain while using 40% less image data. These advancements hold universally across all test scenes, outperforming prior methods in terms of accuracy and completeness.

Remote Sensing Image Change Captioning (RSICC) has emerged as a cross-disciplinary technology that automatically generates sentences describing the changes in bi-temporal remote sensing images. While demonstrating significant potential for urban planning, agricultural surveillance, and disaster management, current RSICC methods exhibit two fundamental limitations: (1) vulnerability to pseudo-changes induced by illumination fluctuations and seasonal transitions and (2) an overemphasis on spatial variations with insufficient modeling of temporal dependencies in multi-temporal contexts. To address these challenges, we present the Frequency–Spatial–Temporal Fusion Network (FST-Net), a novel framework that integrates frequency, spatial, and temporal information for RSICC. Specifically, our Frequency–Spatial Fusion module implements adaptive spectral decomposition to disentangle structural changes from high-frequency noise artifacts, effectively suppressing environmental interference. The Spatia–Temporal Modeling module is further developed to employ state-space guided sequential scanning to capture evolutionary patterns of geospatial changes across temporal dimensions. Additionally, a unified dual-task decoder architecture bridges pixel-level change detection with semantic-level change captioning, achieving joint optimization of localization precision and description accuracy. Experiments on the LEVIR-MCI dataset demonstrate that our FSTNet outperforms previous methods by 3.65% on BLEU-4 and 4.08% on CIDEr-D, establishing new performance standards for RSICC.