Natural language-guided drone geo-localization is essential for enabling autonomous drones to navigate and locate targets in complex environments where GNSS signals are unreliable. However, conventional methods often rely on strict one-to-one matching between images and text, which fails to handle semantic ambiguity—where a single location may have multiple visual perspectives or diverse textual descriptions. To address these challenges, we propose the Dual-Set Contrastive Learning (D-SCL) framework. Unlike prior approaches that represent samples as single points, D-SCL treats images and text as sets to better capture many-to-many semantic associations. The framework introduces two key strategies: Single-Granularity Set Alignment (SG-SA), which uses multi-sample support sets to resolve descriptive ambiguity, and Multi-Granularity Set Alignment (MG-SA), which links local details to global contexts to establish a hierarchical understanding of the scene. Furthermore, we optimize the feature space through intra-modal alignment to enhance the discriminability of visual and textual representations. Extensive experiments on the GeoText-1652 benchmark demonstrate that our method significantly outperforms established baselines, providing a more robust and precise solution for cross-modal drone localization tasks.

Reinforcement Learning (RL) methods typically rely on costly trial-and-error mechanisms to learn decision-making policies from environmental states to agent actions. However, this state-to-action learning paradigm causes low sample efficiency and confines policies to training data distributions, severely limiting their generalization to novel scenes. This challenge is especially critical in long-horizon problems like path planning, which demands robust reasoning across unseen environments. To overcome these limitations, we propose a novel Generalizable Subgoal-Induced Reinforcement Learning (GSIRL) framework. GSIRL fundamentally decouples high-level subgoal decision-making from low-level action execution. Applied to the path planning problem, it learns a generalizable state-to-subgoal policy that adaptively decomposes the global navigation task into a sequence of subgoal-induced tasks. Each generated subgoal is reliably executed by a deterministic executor. Extensive experiments demonstrate that GSIRL achieves high sample efficiency with strong generalization across diverse new scenes.

Few-shot class-incremental learning is crucial for developing scalable and adaptive intelligent systems, as it enables models to acquire new classes with minimal annotated data while safeguarding the previously accumulated knowledge. Nonetheless, existing methods deal with continuous data streams in a centralized manner, limiting their applicability in scenarios that prioritize data privacy and security. To this end, this paper introduces federated few-shot class-incremental learning, a decentralized machine learning paradigm tailored to progressively learn new classes from scarce data distributed across multiple clients. In this learning paradigm, clients locally update their models with new classes while preserving data privacy, and then transmit the model updates to a central server where they are aggregated globally. However, this paradigm faces several issues, such as difficulties in few-shot learning, catastrophic forgetting, and data heterogeneity. To address these challenges, we present a synthetic data-driven framework that leverages replay buffer data to maintain existing knowledge and facilitate the acquisition of new knowledge. Within this framework, a noise-aware generative replay module is developed to fine-tune local models with a balance of new and replay data, while generating synthetic data of new classes to further expand the replay buffer for future tasks. Furthermore, a class-specific weighted aggregation strategy is designed to tackle data heterogeneity by adaptively aggregating class-specific parameters based on local models performance on synthetic data. This enables effective global model optimization without direct access to client data. Comprehensive experiments across three widely-used datasets underscore the effectiveness and preeminence of the introduced framework.

Federated learning aims to collaboratively train a shared global model from multiple clients while preserving data privacy. However, real-world applications often involve clients learning from limited and dynamically arriving data, requiring the global model to classify all encountered classes. This paper introduces federated few-shot class-incremental learning, enabling effective learning of new classes from scarce samples within a decentralized framework. Existing methods suffer from new class bias, where new classes are often misclassified as previously learned ones. Additionally, they face local bias due to non-IID data distribution, which leads client models to focus excessively on their specific local data characteristics. We propose a Decoupled Multi-Prototype Collaboration (DMPC) method to mitigate both biases. First, we introduce a Global Consistency Aggregation mechanism (GCA) that re-weights local prototypes based on their consistency, resulting in more representative global prototypes and effectively eliminating local bias. Second, we design a Multi-Prototype Testing strategy (MPT) that enhances classification accuracy by leveraging both local and global prototypes, thereby mitigating new class bias. More importantly, GCA and MPT exhibit significant synergistic effects. Extensive experiments on three widely used datasets demonstrate the robustness and superiority of our method in bias reduction.

With the rapid advancement of computer vision technology, cross-view geo-localization has become a crucial research focus, especially in UAV-view geo-localization. Existing methods have made significant progress in addressing viewpoint and altitude discrepancies. However, they frequently overlook environmental variability, which can significantly impact both image quality and the resulting localization accuracy. To address these issues, we introduce a novel text-guided environment-invariant feature learning module, which leverages semantic information from textual environment descriptions to guide the learning of robust and environment-invariant feature representations. Furthermore, we employ an enhanced parallel attention mechanism. This mechanism integrates channel attention and pixel-level attention in a parallel structure to simultaneously capture both global features and fine-grained details. Experimental results on the University-1652 dataset demonstrate the effectiveness of the proposed method, confirming its robustness and adaptability in various environments.

This article addresses the unmanned aerial vehicle (UAV)-view geo-localization task, which focuses on bi-directional retrieval between UAV- and satellite-view images. Generally, existing methods aim to learn image representations that can distinguish between different locations while effectively mitigating the cross-view domain gap. However, these methods often struggle in noisy UAV flight environments, as they fail to account for environmental domain shifts caused by varying weather and lighting conditions. To this end, we propose a novel environment-agnostic geo-localization (EAGLe) framework, which integrates a dual-objective discriminator and a style mixture module into diverse UAV-view geo-localization networks to enhance their robustness in dynamic environments. Specifically, the dual-objective discriminator not only distinguishes between UAV and satellite views but also identifies various environmental styles in UAV-view images. Through adversarial learning, the dual-objective discriminator encourages the feature encoder to produce features that remain invariant to both viewpoint and environmental variations. Furthermore, the style mixture module is integrated into the feature encoder to extend diversity at the feature level, allowing EAGLe to learn a broader range of environmental styles beyond the training data. Extensive experiments on the University-1652 and SUES-200 datasets demonstrate that the proposed EAGLe significantly improves the reliability of UAV-view geo-localization networks under dynamic and unpredictable environmental conditions while maintaining inference efficiency.

Skeleton-based action recognition technology has gained significant attention and made great progress in recent years. However, the performance of existing methods declines significantly when the quality of skeleton data extracted by pose estimation algorithms varies. To address this issue, this study proposes an instance-specific learning method aimed at enhancing the model’s ability to learn discriminative features when handling skeleton data of varying quality. We introduce a Dynamic Instance Discriminability Assessment (DIDA) mechanism and a Staged Instance Weighting (SIW) strategy. The DIDA mechanism dynamically evaluates the discriminability of instances by combining prior knowledge with feedback from the model during the training process. The SIW strategy adjusts the weights of instances at different training stages based on their discriminability. Notably, our method requires only a minimal increase in computational cost during training and incurs no additional computational overhead during testing compared to baseline models. We utilized Pifpaf and HR-Net pose estimation methods to extract skeleton data of varying quality from the NTU60, NTU120, and HMDB51 video datasets and conducted extensive experimental validation. The results indicate that the proposed method significantly enhances the action recognition performance while maintaining computational efficiency.

This article explores the problem of skeleton-based one-shot action recognition (SOAR), aiming to build a high-performance action classifier using only a single reference skeleton sequence per action. The key motivation is to develop an embedding space that can be generalized from previously learned, data-abundant actions to novel, data-scarce ones. Nevertheless, this task becomes considerably challenging when the skeleton data are disrupted through occlusions, a common occurrence in real-world applications. This work proposes a novel adaptive part-level embedding graph convolutional network (APLE-GCN), which leverages multistream data streams to extract joint-level features and adaptively aggregate them into part-level embeddings. We go beyond traditional part-based models by not adhering to a fixed strategy of dividing the skeleton into body parts. Instead, we develop a data-driven method to learn instance-specific part templates and compute part-level embeddings based on the correlation between joint-level features and these templates. Unlike body parts defined by skeleton topology, parts extracted via the proposed method, referred to as semantic parts, are intended to capture characteristic motion patterns of various actions and offer increased flexibility when dealing with occlusions. Extensive experiments conducted on five benchmarks demonstrate that the embeddings of semantic parts exhibit strong generalization ability, bringing our model to the state-of-the-art.

In this paper, we tackle the challenge of one-shot action recognition from occluded skeleton data. The task uses limited labeled samples to learn an action recognition model to cope with the time-consuming and laborious problem of annotating a large amount of data. Additionally, the unavoidable occlusion of human joints in real-life scenarios significantly complicates skeleton action recognition efforts. To address these issues, we introduce a novel Spatial-Temporal Denoising AutoEncoder (ST-DAE) designed for skeleton-based one-shot action recognition. Employing an encoder-decoder architecture, the encoder extracts intrinsic patterns of movements from input occluded skeletons, whereas the decoder learns to reconstruct the skeletons under the supervision of clean skeletons. Then a Spatial-Temporal GCN (ST-GCN) backbone network learns to map the reconstructed skeletons into a high-dimensional feature space to capture their essential characteristics for action recognition. Finally, one-shot action recognition is achieved by a KNN classifier in the learned feature space. Experimental validations on the comprehensive NTU-RGB+D 120 RE occlusion dataset demonstrate the effectiveness of the proposed spatial-temporal denoising autoencoder against noisy skeleton data.

In recent years, significant progress has been made in Multi-Agent Deep Reinforcement Learning (MADRL) for addressing cooperative decision-making challenges in multi-aircraft air combat tasks. This paper introduces a novel Spatio-Temporal Relationship Graph Structure Learning method (STRGSL), aimed at overcoming the challenges in capturing the complex and dynamic interactions between agents. The proposed STRGSL constructs a historical behavior graph based on past observations as well as a real-time interaction graph from current observations, providing a comprehensive consideration of both immediate and long-term agent relationships. Leveraging a Graph Neural Network (GNN), STRGSL generates agent representations that fuse current and historical relationships. By integrating STRGSL into a MADRL framework, we jointly optimize both the structure of relationship graphs and the cooperative policies of agents. Experiments carried out in an aircraft combat scenario and two multi-agent cooperative scenarios demonstrate that the proposed STRGSL promotes collaboration among multiple agents, thereby enhancing the overall performance across different scenarios.

This letter aims to deal with the UAV-view geolocalization problem, which is essentially to achieve bi-directional cross-view matching between UAV-view and satellite-view images. The existing studies have confirmed the importance of learning part-wise representations for this task. We go a step further by proposing a two-stage learning framework. The first stage focuses on extracting part-wise representations. In the second stage, a novel Adaptive Embedding Network (AEN) integrates these representations into a global embedding of the entire image to avoid an equal influence of all local parts on image similarity measures. Current mainstream methods typically employ CrossEntropy loss to learn location-dependent representations, aiming to push the distance between different locations in the learned representation space. Some approaches also utilize KL loss or Triplet loss to bring a pair of UAV-satellite images from the same location closer for learning view-invariant representations. However, they overlook a critical concern: a notable representation bias exists among UA-view images captured from the same location but at different viewpoints or heights. To address these issues, we devise a novel cross-view matching loss that narrows the distance between the global embeddings of a satellite-view image and the affinity-aware prototype of multiple true-matched UAV-view images. The experimental results on the University1652 dataset indicate that similarity measures in the learned embedding space exhibit excellent generalization to images from new locations, achieving superior cross-view matching performance compared to previous methods

Unmanned Aerial Vehicle (UAV) visual geo-localization aims to match images of the same geographic target captured from different views, i.e., the UAV view and the satellite view. It is very challenging due to the large appearance differences in UAV-satellite image pairs. Previous works map images captured by UAVs and satellites to a shared feature space and employ a classification framework to learn location-dependent features while neglecting the overall distribution shift between the UAV view and the satellite view. In this paper, we address these limitations by introducing distribution alignment of the two views to shorten their distance in a common space. Specifically, we propose an end-to-end network, called PVDA (Progressive View Distribution Alignment). During training, feature encoder, location classifier, and view discriminator are jointly optimized by a novel progressive adversarial learning strategy. Competition between feature encoder and view discriminator prompts both of them to be stronger. It turns out that the adversarial learning is progressively emphasized until UAV-view images are indistinguishable from satellite-view images. As a result, the proposed PVDA becomes powerful in learning location-dependent yet view-invariant features with good scalability towards unseen images of new locations. Compared to the state-of-the-art methods, the proposed PVDA requires less inference time but has achieved superior performance on the University-1652 dataset.

This paper addresses the task of Unmanned Aerial Vehicles (UAV) visual geo-localization, which aims to match images of the same geographic target taken by different platforms, i.e., UAVs and satellites. In general, the key to achieving accurate UAV-satellite image matching lies in extracting visual features that are robust against viewpoint changes, scale variations, and rotations. Current works have shown that part matching is crucial for UAV visual geo-localization since part-level representations can capture image details and help to understand the semantic information of scenes. However, the importance of preserving semantic characteristics in part-level representations is not well discussed. In this paper, we introduce a transformer-based adaptive semantic aggregation method that regards parts as the most representative semantics in an image. Correlations of image patches to different parts are learned in terms of the transformer's feature map. Then our method decomposes part-level features into an adaptive sum of all patch features. By doing this, the learned parts are encouraged to focus on patches with typical semantics. Extensive experiments on the University-1652 dataset have shown the superiority of our method over the current works.

State-of-art statistical models for object recognition are sensitive to large affine transformations in scale and rotation. This is due to the rather strict spacial correlation assumption made between model parts that cannot be solved using distance transform. In this work, to tackle these problems we take a different approach by utilizing topological properties of a graphical object model. Our model is a part-based generative model with each part representing a patch of an object image. The graph structure is learned by Boosting the graph nodes additively so that new parts are added in until a threshold is reached. Every time when a new part is inserted, all other parts are updated accordingly. The result of this process gives us a reliable and flexible graph structure that encodes local topological information of an object appearance. The recognition step is performed using two different methods: either by exhaustive search of a topological graph which best matches the model, or by fitting the model incrementally starting from selecting the most reliable (rather than most dominant from training) part candidates from the test image.

This paper exploits the properties of the commute time for the purposes of graph simplification and matching. Our starting point is the lazy random walk on the graph, which is determined by the heat kernel of the graph and can be computed from the spectrum of the graph Laplacian. We characterise the random walk using the commute time between nodes, and show how this quantity may be computed from the Laplacian spectrum using the discrete Green's function. In this paper, we explore two different, but essentially dual, simplified graph representations delivered by the commute time. The first representation decomposes graphs into concentric layers. To do this we augment the graph with an auxiliary node which acts as a heat source. We use the pattern of commute times from this node to decompose the graph into a sequence of layers. Our second representation is based on the minimum spanning tree of the commute time matrix. The spanning trees located using commute time prove to be stable to structural variations. We match the graphs by applying a tree-matching method to the spanning trees. We experiment with the method on synthetic and real-world image data, where it proves to be effective.

This paper explores the use of commute-time preserving embedding as means of data-clustering. Commute time is a measure of the time taken for a random walk to set-out and return between a pair of nodes on a graph. It may be computed from the spectrum of the Laplacian matrix. Since the commute time is averaged over all potential paths between a pair of nodes, it is potentially robust to variations in graph structure due to edge insertions or deletions. Here we demonstrate how nodes of a graph can be embedded in a vector space in a manner that preserves commute time. We present a number of important properties of the embedding. We experiment with the method for separating object motions in image sequences.

The presence of noise renders the classical factorization method almost impractical for real-world multi-body motion tracking problems. The main problem stems from the effect of noise on the shape interaction matrix, which looses its block-diagonal structure and as a result the assignment of elements to objects becomes difficult. The aim in this paper is to overcome this problem using graph-spectral embedding and the k-means algorithm. To this end we develop a representation based on the commute time between nodes on a graph. The commute time (i.e. the expected time taken for a random walk to travel between two nodes and return) can be computed from the Laplacian spectrum using the discrete Green’s function, and is an important property of the random walk on a graph. The commute time is a more robust measure of the proximity of data than the raw proximity matrix. Our embedding procedure preserves commute time, and is closely akin to kernel PCA, the Laplacian eigenmap and the diffusion map. We illustrate the results both on the synthetic image sequences and real world video sequences, and compare our results with several alternative methods.

This paper exploits the properties of the commute time for the purposes of graph matching. Our starting point is the lazy random walk on the graph, which is determined by the heat-kernel of the graph and can be computed from the spectrum of the graph Laplacian. We characterise the random walk using the commute time between nodes, and show how this quantity may be computed from the Laplacian spectrum using the discrete Green’s function. We use the commute-time to locate the minimum spanning tree of the graph. The spanning trees located using commute time prove to be stable to structural variations. We match the graphs by applying a tree-matching method to the spanning trees. We experiment with the method on synthetic and real-world image data, where it proves to to be effective.

This paper exploits the properties of the commute time for the purposes of graph matching. Our starting point is the random walk on the graph, which is determined by the heat-kernel of the graph and can be computed from the spectrum of the graph Laplacian. We characterise the random walk using the commute time between nodes, and show how this quantity may be computed from the Laplacian spectrum using the discrete Green's function. We use the commute-time to locate the minimum spanning tree of the graph. The spanning trees located using commute time prove to be stable to structural variations. We match the graphs by applying a tree-matching method to the spanning trees. We experiment with the method on synthetic and real-world image data, where it proves to be effective

Although inexact graph-matching is a problem of potentially exponential complexity, the problem may be simplified by decomposing the graphs to be matched into smaller subgraphs. If this is done, then the process may cast into a hierarchical framework and hence rendered suitable for parallel computation. In this paper we describe a spectral method which can be used to partition graphs into non-overlapping subgraphs. In particular, we demonstrate how the Fiedler-vector of the Laplacian matrix can be used to decompose graphs into non-overlapping neighbourhoods that can be used for the purposes of both matching and clustering.