Online map generation from remote sensing (RS) imagery represents a critical yet challenging task in the geographic information system. With the development of deep learning, automatic generation has become a popular research focus. Current mainstream approaches involve two paradigms: domain mapping utilizing generative models and semantic segmentation based on discriminative models. Although these methods provide an effective tool for automatic online map generation, they still face a fundamental dilemma of suffering from poor inferential creativity or low semantic consistency. Therefore, we argue that lacking proper guidance may result in some semantic distortion in the generated maps. To address this issue, we propose PSP-MapDiff, a novel end-to-end network based on a generative denoising diffusion model (DM) incorporating semantic prior information provided by pretrained semantic segmentation networks. In particular, we design a conditional diffusion framework with a denoising starting point from source domain images, effectively driving the denoising process in a promising direction. In addition, the proposed semantic consistency gradient (SCG) and boundary preservation gradient (BPG) modules progressively utilize semantic prior information to enhance cross-domain translation accuracy. Geometric consistency loss is further designed to mitigate distortion and improve robustness. Extensive experiments demonstrate that PSP-MapDiff achieves improvements in visual quality (e.g., +27.8% in LPIPS) and accuracy (e.g., +2.3% in ACC) compared to other state-of-the-art methods, indicating the superiority of the proposed model in maintaining the semantic details of the source domain images.

In affine formation control problems, the construction of the framework with universal rigidity and affine localizability is a critical prerequisite, but it has not yet been well addressed, especially when additional agents join the formation or link/agent failures emerge. Motivated by this observation, we investigate the problem of constructing affine formation frameworks in three scenarios, including vertex addition, edge deletion and vertex deletion. Our approach starts from the original affine formation and uses geometric methods to locally adjust the structure of the weighted graph to describe the topology, so that the modified framework maintains the universal rigidity and affine localizability. Notably, the developed strategies only utilize local measurements and exhibit distributed characteristics, laying the foundation for applications in multi-agent systems. To demonstrate the compatibility with affine formation control proposals, we present a case study on affine formation tracking in a multi-UAV formation, demonstrating the effectiveness of our algorithms in constructing eligible frameworks in aforementioned scenarios. Moreover, a comparative simulations is also conducted to highlight the low time complexity of our distributed algorithm relative to the centralized optimization-based method.

Road extraction from remote sensing images is crucial for autonomous driving, urban planning, and disaster response, yet remains challenging due to the presence of complex interference, the elongated structure of roads, and occlusions by surrounding objects. To tackle these challenges, we propose SAM4RoadEx, a dedicated architecture for road extraction from remote sensing images. First, the encoder incorporates the pretrained knowledge of SAM to effectively model road features and suppress background noise. Second, a structure-aware guided decoder is specifically designed to capture elongated road structures, enabling accurate, and complete road extraction. In particular, it first applies cross-attention between learnable road tokens and image features to capture global structure. Then, it uses self-attention to refine spatial continuity. Finally, a prompt-guided active supervision (PGAS) strategy is designed to provide guidance in severely occluded or discontinuous predictions. During training, PGAS evaluates prediction confidence and, when necessary, dynamically generates point prompts in uncertain regions using distance-transform assessment to guide iterative refinement. These prompts guide the network to focus on discontinuous areas and iteratively refine its predictions. Experiments on the widely used DeepGlobe Road, Massachusetts Road, and CHN6-CUG datasets show that our method outperforms several state-of-the-art methods, achieving an intersection over union increase of 1.03%-4.74% and an average path length similarity increase of 1.22%-5.60%. SAM4RoadEx achieves superior performance over several state-of-the-art methods.

Online map plays a significant role in modern society. Especially with the rapid expansion of urbanization, obtaining updated maps in a timely manner has become a necessary task. Recently, the rise of generative models has provided an alternative tool for efficient map generation. However, existing methods focus on single-domain translation from remote sensing (RS) images to online maps, which may suffer from poor generalization. For remote sensing images acquired from different regional domains with significant difference, a model trained on one domain can hardly be employed for images in other domains. To address this issue, we propose a novel cross-domain adaptation framework for online map generation from different areas. Our framework comprises two stages. In the first stage, we propose a cross-domain generative adversarial network for source-to-target domain RS images translation, applying the feature-consistency and geometric-consistency constraints. We acquire source-domain RS images with the style of the target domain. In the second stage, we design a conditional diffusion model for online maps generation from translated RS images. To guide the denoising process to a promising direction, we design a content prompt encoder to facilitate the semantic transfer of RS images to the generated online maps. Extensive experiments are conducted in cross-domain datasets. The results demonstrate the superior performance of our model over the state-of-the-art methods on the cross-domain online map generation.

In recent years, vehicle detection in uncrewed aerial vehicle (UAV) videos has attracted significant attention. However, the traditional detection paradigm based on single frame image is faced with the performance bottleneck caused by rare pose, motion blur, and occlusion. To address these issues, we propose a novel network comprising three important modules designed to fully exploit the spatiotemporal contextual information across video frames. First, To strengthen the backbone network’s capacity for extracting global contextual information from single-frame images, we propose the visual state space context module (VSSCM). By incorporating the 2D-selective-scan module (SS2D), VSSCM captures global dependencies and enriches contextual information without significantly increasing computational complexity. Second, the temporal information-guided spatial attention aggregation module (TGSAM) is introduced to fuse features from critical regions in adjacent frames. Finally, the self-attention-based classification feature aggregation module (SACFAM) is designed to model the relation between object features across frames and perform feature aggregation based on the learned relation matrix, thereby effectively improving the quality of classification features in UAV videos. Extensive experiments are conducted on the challenging VisDrone2019-VID dataset, and the experimental results demonstrate the effectiveness and superiority of the proposed method.

Cross-View Object Geo-Localization (CVOGL) aims to accurately locate specified objects within ground or drone-view images onto their corresponding positions in satellite imagery, addressing a critical demand for fine-grained geo-localization beyond coarse scene-level estimation. This fine-grained localization capability plays a vital role in various practical fields, such as security surveillance, autonomous driving, smart map construction, and robotics. However, CVOGL faces significant challenges due to severe viewpoint differences, data scarcity, and the difficulty of establishing robust cross-view object correspondences. Inspired by the observation that view-invariant properties are shared between scene-level and object-level features, we propose STONet, a novel cross-task knowledge transfer framework for CVOGL. STONet employs a two-stage training pipeline. In the first stage, it first leverages Cross-View Image Geo-Localization (CVIGL) to learn foundational scene-level view-invariant features for the backbone. In the second stage, a cross-task knowledge transfer mechanism is designed to adapt this scene-level knowledge for object-level geo-localization. To ensure robust learning and prevent catastrophic forgetting, we introduce the Object-Scene Joint Loss, which balances global scene-level constraints with fine-grained object-level supervision. Additionally, to address feature space misalignment caused by the orthogonal view difference between ground panoramic images and satellite images, we propose a geometry-constrained inverse-polar feature transform for ground features. On the CVOGL dataset, STONet outperformed state-of-the-art performance with only 29.52M parameters (↓60%) and 105.39 GFlops (↓50%).

Remote sensing instance segmentation plays a vital role in geographic information systems, and it is crucial for Internet of Things applications such as smart city infrastructure management, traffic monitoring, and autonomous driving systems. Although fully supervised methods have achieved promising accuracy, they rely heavily on large amounts of pixel-level annotations, which are extremely costly to obtain for high-resolution remote sensing images. Box-supervised instance segmentation typically leverages horizontal bounding boxes as weak supervision signals, significantly reducing the annotation burden. However, mask prediction under box supervision faces challenges due to limited utilization of spatial information, including the lack of geometric details, the disconnect between spatial localization and pixel-level prediction, and the neglect of spatial priors in traditional pairwise affinity loss. To address these issues, this article proposes SALInst, a spatial affinity learning framework for box-supervised remote sensing instance segmentation. Specifically, a spatial information enhancement module and a dual-stream residual gate fusion mechanism are designed to strengthen spatial awareness and semantic coherence of mask features. Furthermore, by leveraging the spatial constraint prior, we propose a spatial affinity loss based on Gaussian kernel and total variation loss to reinforce the spatial consistency of predicted masks. Extensive experiments on the iSAID and NWPU VHR-10 datasets demonstrate that SALInst outperforms existing box-supervised methods while narrowing the performance gap between weakly and fully supervised instance segmentation.

Rapid 3D building reconstruction in urban-scale areas has emerged as a pivotal technology for smart city applications. Recent methods that reconstruct buildings from single off-nadir imagery have gained attention due to their efficiency in both time and data costs. However, the training of these methods relies on large-scale, costly 3D annotations, including building bounding boxes, roofs, footprints, and roof-to-footprint offsets, and thus cannot be trained when only the footprint is available, despite the fact that a large amount of building footprints can be easily obtained in crowdsourced building data set form the Internet. To address this, we propose a semi-weakly supervised learning method that leverages massive weakly annotated data (footprints) and a limited number of manually annotated 3D building labels to learn to reconstruct 3D buildings. In our method, we introduce an ingenious wireframe representation to replace conventional bounding-box representation, thereby providing a foundation for semi-weakly supervised learning. Based on this representation, we propose BWFNet for extracting building wireframes. BWFNet enhances accuracy under semi-weakly supervision by modeling both structural and local knowledge. Furthermore, we propose a training strategy for building wireframe extraction grounded in the principle of geometric consistency constraints to further improve semi-weakly supervised learning. The experimental results demonstrate that the proposed BWFNet achieves excellent reconstruction performance by requiring full annotation for only 3% of the training data. This performance represents a significant improvement compared to current state-of-the-art methods. The code for BWFNet is open-sourced at https://shaoruizhe.github.io/BWFNet.github.io/.

Medical image analysis is a cornerstone of modern healthcare, yet conventional single-modal deep learning often struggles with the unique physical constraints and structural variability inherent in data acquired from diverse medical sensors. Recently, Vision-Language Models (VLMs) have sparked a paradigm shift by bridging the semantic gap between visual sensor signals and clinical narratives. Following the PRISMA guidelines, 167 representative studies are systematically synthesized in this review to provide a comprehensive roadmap of VLM technological evolution and clinical utility. First, rather than treating VLMs as generic feature extractors, their underlying mechanisms are uniquely distilled into seven core operational principles, which are then explicitly mapped to downstream applications such as few-shot diagnosis, prompt-driven segmentation, and multi-task foundation models. To facilitate intuitive evaluation, a rigorous quantitative cross-comparison of current benchmark architectures is presented. Crucially, this review goes beyond highlighting successes by critically assessing prevalent clinical bottlenecks, including zero-shot segmentation failures, multi-modal hallucinations in diagnosing rare diseases, and the prohibitive computational complexity associated with 3D volumes and gigapixel whole slide images. Finally, a novel, forward-looking framework is proposed: the transition from static “image-text alignment” to dynamic “multi-source sensor-driven intelligence”. By addressing both physical sensor constraints and algorithmic limitations, this survey offers actionable insights for developing trustworthy, sensor-aware clinical diagnostic agents.

Points of Interest (POIs) data are vital for location-based services, yet their production remains challenging due to labor-intensive collection and verification processes. Generating POIs from street-view imagery (SVI) has recently emerged as a promising solution. However, the lack of open benchmark hinders its development. Existing methods typically treat SVI as isolated images without fully leverage their multi-view representations of geographical entities. We present SVI2POI, a novel end-to-end framework for POI extraction from SVI. It brings two key innovations. In the signboard recognition stage, the proposed YOLOv11s-DLKA detector enhanced performance degraded by geometric distortions commonly occurred in SVI. In the POI generation stage, we propose a clustering strategy combined with large language model-based naming and photogrammetric positioning to consolidate multi-view information for accurate POI identification. Furthermore, we introduce the first open dataset for end-to-end POI generation from SVI. It contains a training dataset including 3,097 SVIs with 13,182 manually annotated regions of interest (ROIs), and a benchmark dataset with 927 SVIs and 1,004 manually labeled POIs, with 190 verified against OpenStreetMap-POIs and therefore contains coordinates. Our framework achieves 61.69% precision, 50.70% recall, and 55.66% F1-score, outperforms state-of-art method with 5.59%, 1.16%, and 3.04%, respectively, via cross-method and cross-dataset comparison.

Existing feed-forward 3D reconstruction methods typically process a bounded number of images and recover cameras and geometry in local or internally normalized frames. Extending them to large-scale UAV mapping requires scalable multi-chunk processing and reliable aggregation, while full Sim(3) alignment can become unstable for near collinear trajectories. We present GeoFF3D, which combines a coordinate-anchored model with a spatial large-scale reconstruction framework (SLRF). The model uses georeferenced camera translations and optional geometric priors to predict camera poses and dense point maps directly in a gravity-aligned Z-up metric frame. SLRF partitions images into spatially overlapping chunks, propagates shared-view priors, and aggregates local reconstructions hierarchically, while remaining applicable to different bounded-view models. Across nine aerial mapping blocks, GeoFF3D achieves the best average reconstruction quality, improving F@5 from 0.829 for Pi3X + SLRF to 0.877. On long UAVScenes sequences, it reaches 0.848, compared with 0.687 for Pi3X + SLRF and 0.451 for the strongest evaluated SLAM/streaming baseline. GeoFF3D reconstructs 2,000 images in approximately five minutes, demonstrating scalable and robust large-scale UAV reconstruction.The code is available at https://github.com/yanxian-ll/GeoFF3D.

Image fusion aims to synthesize comprehensive scene representations by integrating complementary characteristics from multi-source images. Inspired from the human knowledge prior, existing methods typically focus on extracting predefined features to assist the fusion process, e.g., texture structure in visible images and thermal radiation in infrared images. However, such task-dependent schemes inevitably neglect the rich diversity of latent characteristics in multi-source image, which not only limits the exploration of the rich complementary cross-source information, but also degrades the generalization capability to diverse fusion tasks. In this paper, we propose a simple and unified method for generalized image fusion. Specifically, we generalize the image fusion task into common information collection and unique information maintenance, and design a common-unique decoupling network with a tri-encoder structure, in which a shared encoder and two private encoders are developed to extract common and unique features, respectively. To achieve robust common-unique information disentanglement, we design a decoupling loss that constrains the mutual-information among the tri-encoder outputs and source features. Unlike conventional approaches requiring human knowledge priors, our method automatically learns essential characteristics directly from input images. We conduct comprehensive experiments covering four image fusion tasks on nine popular datasets. Extensive experimental results demonstrate that our method achieves superior performance on all these tasks as compared of the respective state-of-the-art fusion methods, and show impressive cross-task generalization capability.

Monocular depth estimation (MDE) is a cornerstone task in 2D/3D scene reconstruction and recognition with widespread applications in autonomous driving, robotics, and augmented reality. However, existing state-of-the-art methods face a fundamental trade-off between computational efficiency and estimation accuracy, limiting their deployment in resource-constrained real-world scenarios. It is of high interest to design lightweight but effective models to enable potential deployment on resource-constrained mobile devices. To address this problem, we present RepACNet, a novel lightweight network that addresses this challenge through reparameterized asymmetric convolution designs and CNN-based architecture that integrates MLP-Mixer components. First, we propose Reparameterized Token Mixer with Asymmetric Convolution (RepTMAC), an efficient block that captures long-range dependencies while maintaining linear computational complexity. Unlike Transformer-based methods, our approach achieves global feature interaction with tiny overhead. Second, we introduce Squeeze-and-Excitation Consecutive Dilated Convolutions (SECDCs), which integrates adaptive channel attention with dilated convolutions to capture depth-specific features across multiple scales. We validate the effectiveness of our approach through extensive experiments on two widely recognized benchmarks, NYU Depth v2 and KITTI Eigen. The experimental results demonstrate that our model achieves competitive performance while maintaining significantly fewer parameters compared to state-of-the-art models.

Road extraction from remote sensing (RS) imagery is crucial for urban management, traffic planning, and autonomous driving. However, extracting accurate and complete roads remains challenging due to occlusions and severe class imbalance, where non-road regions dominate. To address these challenges, we propose a novel road extraction method incorporating two key components. The first is a Focused Masked Image Modeling (FocusMIM) strategy for data augmentation, which randomly masks road-related regions to efficiently model the latent dependency between occluded and non-occluded road parts. With FocusMIM, the model's ability to infer occluded roads is obviously improved. The second is a Focused Transformer (FocusFormer), which enhances road-related feature interactions through a Transformer-based encoder with Channel Self-Attention (CSA) modules and a Transformer decoder that leverages masked attention. The CSA modules aggregate global features of RS images to enhance contextual inference and mitigate occlusions. Meanwhile, the Transformer decoder employs a single road query that attends exclusively to road features, alleviating the class imbalance issue. Comprehensive experiments on the DeepGlobe Road, Massachusetts Road, and CHN6-CUG datasets demonstrate that our method outperforms several state-of-the-art methods, achieving an IoU increase of 0.96-5.38%. These results confirm the effectiveness of FocusMIM and FocusFormer in improving road continuity and reducing background interference.

This paper studies the problem of autonomous exploration in unknown environments with a mobile robot platform. By integrating a simultaneous localization and mapping system, the detected boundary points are filtered and used as navigation target points to accomplish autonomous exploration. A novel algorithm for boundary point detection, filtering, and target point assignment is proposed. For boundary point detection, due to the non-completeness of sensor-based random tree methods and the low efficiency of traditional rapidly-exploring random tree methods, this paper employs the probabilistically complete dualthread rapidly-exploring random trees algorithm. The dual-thread rapidly-exploring random trees algorithm adopts a dual-thread architecture that combines global and local trees to improve exploration efficiency while ensuring full coverage of the unknown environment. For boundary point filtering, the K-means clustering algorithm is replaced with the mean shift algorithm, reducing dependency on initial inputs and enhancing algorithm robustness. Regarding target point assignment, a Bayesian decision-making approach based on multiple prior inputs is utilized, ensuring the efficiency of the exploration process. Physical experiments demonstrate that the proposed approach achieves significant performance in the studied scenario.

Recent advances in remote sensing technology have made sub-meter resolution imagery increasingly accessible, offering remarkable detail for fine-grained land cover analysis. However, its full potential remains underutilized—particularly for large-scale land cover mapping—due to the lack of sufficient, high-quality annotated datasets. Existing labels are typically derived from pre-existing products or manual annotation, which are often unreliable or prohibitively expensive, particularly given the rich visual detail and massive data volumes of sub-meter imagery. Inspired by the spatial autocorrelation principle, which suggests that objects of the same class tend to co-occur with similar visual features in local neighborhoods, we propose the Mask Clustering-based Annotation Engine (MCAE), which treats semantically consistent mask groups as the minimal annotating units to enable efficient, simultaneous annotation of multiple instances. It significantly improves annotation efficiency by one to two orders of magnitude, while preserving label quality, semantic diversity, and spatial representativeness. With MCAE, we build a high-quality annotated dataset of about 14 billion labeled pixels, referred to as HiCity-LC, which supports the generation of city-scale land cover maps across five major Chinese cities with classification accuracies above 85%. It is the first publicly available sub-meter resolution city-level land cover benchmark, highlighting the scalability and practical utility of MCAE for large-scale, sub-meter resolution mapping. The dataset is available at https://github.com/chenhaocs/MCAE.

Standoff tracking of a ground-moving target with a single fixed-wing unmanned aerial vehicle (UAV) is vulnerable to occlusions around the target, such as buildings and terrain, which can obstruct the line of sight (LOS) between the UAV and the target, resulting in tracking failures. To address these challenges, multi-UAV cooperative tracking is often employed, offering multi-angle coverage and mitigating the limitations of a single UAV by maintaining continuous target visibility, even when one UAV’s LOS is obstructed. Building on this idea, we propose a specialized two-layer framework for dual-UAV cooperative target tracking. This framework comprises a decision-making layer and a guidance layer. The decision-making layer employs a state-transition-based distributed role transition algorithm for dual UAVs. Here, the UAVs periodically share state variables based on their target observability. In the guidance layer, we devise a velocity-vector-field-based controller to simplify the complexity of controller design for cooperative tracking. To validate the proposed framework, three numerical simulations and one hardware-in-the-loop (HIL) simulation were conducted. These simulations confirmed that the role transition algorithm functions properly even under occlusion conditions. Additionally, the standoff tracking guidance controller demonstrated superior performance compared to baseline methods in terms of tracking accuracy and stability.

High-quality video semantic segmentation has made great progress thanks to advanced deep neural networks. However, two main challenges persist: 1) achieving high-precision segmentation without excessive computational costs, and 2) ensuring semantic consistency in segmentation results. Recent studies have predominantly employed self-attention mechanisms to merge similar features across multiple frames. This fusion typically occurs after deep feature extraction. However, this approach overlooks similar features within individual frames and suffers from information loss during extraction, leading to insufficient fusion of similar features. This, in turn, impacts segmentation accuracy and causes semantic inconsistencies. To address these issues, we propose a novel neural network with Multi-level Global-Local Context Aggregation (MGLA) in this paper. MGLA models similar features both across multiple frames and within a single frame as global information, thus ensuring semantic consistency in the segmentation results. Moreover, MGLA incorporates multi-level multi-frame feature information in the fusion process, complementing each other to prevent information loss and ensuring segmentation accuracy. The proposed method has achieved state-of-the-art results on two challenging public video semantic segmentation datasets: 78.9

Semi-supervised object detection (SSOD) provides a promising solution to mitigate the annotation costs in remote sensing applications. Mainstream teacher-student based SSOD methods leverage unlabeled images through pseudo labeling, and their effectiveness is fundamentally limited by the inevitable noise in pseudo labels, particularly for remote sensing (RS) scenarios with complex backgrounds and dense, multi-scale and oriented objects. Current methods primarily focus on reducing pseudo label noise through category, scale and Intersection over Union information mining, as well as designing fine-grained confidence thresholding strategies. However, the inherent discrepancy between classification and localization reliability is neglected. In this study, with analyzing the characteristic discrepancies between the classification and localization branches, We propose artificial intelligence (AI) methodological innovation method named cross-branch information incorporation method (i.e., CBI-SSOD) to utilize these discrepancies to assist the training of the classification branch, and thus improve the performance of SSOD methods. Specifically, our method present two key AI innovations. Firstly, we propose a pretext task to extract cross-branch information, which can improve the classification ability by reinforce the consistent predictions between the classification branch and the pretext task. Besides, we propose a pseudo label reassignment approach to adjust the soft classification pseudo labels, and thus suppress pseudo label noise and improve the detection performance. Extensive experiments on Dataset for Object Detection in Aerial Images (DOTAv1.0) and DOTAv1.5 datasets validate the effectiveness and superiority of our method, and demonstrate the practical engineering impact of our method on RS applications and interpretation systems.

Mesh is one of the most commonly utilized data formats for digital three-dimensional models in most existing 3-D applications. Recently, online mesh reconstruction from light detection and ranging (LiDAR) measurements has garnered significant interest because of its high efficiency. However, due to the lack of adaptability in adjusting vertex density, existing methods tend to generate either over-represented planar mesh or under-represented non-planar mesh. To address this issue, we propose a novel online mesh reconstruction method with a self-adaptive strategy which, respectively, processes planar and non-planar regions according to local geometric features. For planar regions, we propose a two-step points decimation and mesh reconstruction algorithm to reduce data redundancy based on the observation that the geometric structure of these regions is simple and can be represented by a few key vertices and triangles. For non-planar regions, we design a parallel direct meshing (PDM) algorithm with hole filling mechanism to model objects with complex geometric structure. Moreover, we propose a zipper-based connection strategy to handle the boundaries between planar and non-planar mesh regions. Experimental results demonstrate that our approach outperforms several state-of-the-art algorithms in terms of mesh quality and memory consumption. Remarkably, the entire process is capable of running in real-time on a standard desktop CPU. Code is available at https://github.com/Neo-cyber-hubb/LGFaware-Meshing.