In recent years, the emergence of large visionlanguage models (VLMs) has catalyzed the development of prompt learning, where networks are trained to enhance VLM performance by learning continuous prompts. However, traditional continuous prompt learning often struggles with challenges like overfitting to Base classes and a lack of interpretability due to the nature of prompt parameterization. To overcome these limitations, we introduce Discrete Prompt Optimization (DPO), a method that optimizes text prompts in discrete word-space. During training, scores are assigned to token embeddings, which are then used to select the most effective token sequence for the downstream task. DPO was tested across 11 diverse datasets, consistently outperforming baseline methods like CLIP and CoOp on Novel classes in most cases. This discrete approach not only reduces overfitting but also enhances transparency and model interpretability, enabling the learning of dataset-specific text prompts that are easily understandable.

Scene Graph Generation (SGG) aims to represent visual scenes as structured graphs, where objects and their pairwise relationships are modeled as relational triples. However, conventional SGG methods often struggle with two crucial limitations: inadequate modeling of rich spatial relations beyond 2D layouts, and poor generalization in open-vocabulary scenarios. To address these limitations, we propose SPIN-SGG, a novel framework that explicitly integrates multi-dimensional spatial information—including planar geometry, relative depth, and topological structure—into an open-vocabulary SGG pipeline, without requiring ground-truth 3D annotations. Our approach builds upon a two-stage design. First, we generate pseudo-3D scene reconstructions from monocular images using multi-view synthesis and point cloud estimation, and construct a spatially enriched instruction-tuning dataset that includes fine-grained spatial predicates (e.g., left of, above, inside). Second, we propose a spatially aware visual-language model, trained with both static scene graph descriptions and dynamic spatial reasoning tasks such as question answering and multi-turn dialogue. To further enhance spatial layout consistency, we incorporate layer-aware clustering and object-level depth anchoring in the scene parsing module. Extensive experiments on the PSG benchmark and a newly curated SpatialSGG dataset demonstrate that SPIN-SGG significantly outperforms previous open-vocabulary SGG methods, with improvements of +1.3% in mR@50 on PSG and +2.7% QA accuracy on SpatialSGG, showcasing its robust and comprehensive spatial reasoning capabilities.

LiDAR-based global localization plays a crucial role in autonomous driving and robotic systems. However, real-time changes in environmental conditions present challenges for global localization. Current methods predominantly rely on sparse LiDAR point clouds with single-feature information (semantic or geometric), leading to insufficient accuracy and robustness in 6-DOF pose estimation and similarity matching. In this work, we propose a novel semantic-geometric combined 3D histogram descriptor, called the SGHD descriptor, with robust place recognition and 6-DoF pose estimation capabilities. We first extract semantic and geometric features of objects through operations such as point cloud clustering and semantic segmentation. Next, we describe the distribution of neighboring objects using topological structures. To improve object discriminability, we construct triangular descriptors for each object based on semantic information and relative distance. Subsequently, we consider the complementary nature of semantic and geometric information and propose a 3D histogram descriptor. This 3D histogram descriptor encodes semantic, spatial angle and relative distance information, enhancing the invariance of the triangle descriptor. We perform similarity discrimination and 6-DOF pose estimation through coarse-to-fine matching of local and global semantic maps. Overall, the proposed descriptor achieves high-accuracy and high-robustness 6-DOF pose estimation and place recognition for global localization. We extensively compared the proposed SGHD descriptor with state-of-the-art methods on the public SemanticKITTI dataset. Quantitative results demonstrate that SGHD exhibits higher adaptability and significant improvements compared to its counterparts. The implementation of SGHD will be released at https://github.com/wf-hahaha/SGHD

Low Earth orbit (LEO) satellites are increasingly required to process and securely stream video data for remote sensing, surveillance, and onboard perception applications. However, the strict constraints of onboard computing capability, power budget, and thermal dissipation make conventional encryption schemes difficult to apply to real-time video streaming tasks. To address this challenge, this paper proposes a low-complexity real-time video encryption algorithm for resource-constrained LEO satellites. The proposed method integrates selective encryption with a lightweight permutation–diffusion mechanism to reduce computational overhead while maintaining effective protection of continuous video streams. To enhance security, a chaotic pseudo-random sequence generator is employed to improve encryption randomness, and a dynamic key scheduling strategy is introduced to increase temporal key variability and strengthen resistance to statistical and differential attacks across successive frames. The algorithm is further designed for efficient deployment on embedded onboard platforms with limited hardware resources. Experimental results show that the proposed method achieves favorable performance in encryption speed, computational complexity, information entropy, adjacent pixel correlation, and differential attack resistance. Compared with conventional full-encryption methods, the proposed algorithm offers a more balanced trade-off between security and real-time efficiency, demonstrating its potential for secure video streaming in resource-constrained LEO satellite systems.

Ground-based optical remote sensing of aerial targets at kilometer-scale standoff distances requires accurate keypoint localization for six-degree-of-freedom (6-DoF) pose recovery under variable illumination, motion blur, and atmospheric degradation. Many lightweight detectors use fixed-kernel convolutions, whose spatially invariant sampling may limit adaptation to heterogeneous target geometries and spatially varying image degradation. We introduce GeoAdapt, a compact keypoint detection framework that inserts deformable convolution v2 (DCNv2) modules between the feature pyramid network (FPN) neck and the detection head. GeoAdapt also replaces the standard object keypoint similarity (OKS) loss with a combination of Wing Loss and Bone Loss. The complete model contains 5.95 M parameters, 47.9% fewer than the You Only Look Once version 8 small pose model (YOLOv8s-pose). On a synthetic ground-based optical remote sensing benchmark, GeoAdapt achieved a percentage of correct keypoints (PCK) at a threshold of 0.05 times the bounding-box diagonal (PCK@0.05D) of 89.3% and a rotation error of 11.6°, improving PCK by 11.7 percentage points over YOLOv8s-pose. Zero-shot evaluation on manually annotated real ScanEagle and Matrice 200 imagery showed consistent advantages over YOLOv8s-pose and YOLO11s-pose in all six test scenarios. A factorial ablation indicated a positive interaction between DCNv2 and the Wing+Bone loss.

Adverse weather poses a major challenge to ground-based electro-optical (EO) aircraft surveillance. Existing parameter-efficient fine-tuning (PEFT) methods operate mainly in the spatial domain and treat weather as a generic domain shift rather than a frequency-dependent degradation. We propose Freq-LoRA, a frequency-domain PEFT method that applies the Type II Discrete Cosine Transform (DCT-II) to frozen encoder features, decomposes them into $K=4$ learned Gaussian frequency bands, and modulates the bands with an image-driven spectral gate. SpectralGate contains 140 parameters and estimates band importance from the input image’s DCT statistics, removing the need for external weather metadata at inference; weather-diverse training data are still required to learn the decomposition. On a Blender-simulated aircraft dataset covering five weather conditions, Freq-LoRA achieved a test mean Intersection-over-Union (mIoU) of 0.904, with a 95% confidence interval (CI) of [0.899, 0.908], using 559 K trainable parameters. Its point estimate differed by $-0.002$ from that of the weather-conditioned spatial method Feature-wise Linear Modulation (FiLM; 0.906), despite requiring no external weather metadata at inference. Relative to batch-size-matched Spatial LoRA (0.873; 95% CI: [0.867, 0.879]), Freq-LoRA had a 0.031 higher point estimate. Preliminary evaluation on real unmanned aerial vehicle (UAV) imagery yielded 0.421 mIoU (+13% relative to the zero-shot Segment Anything Model (SAM); one platform), and evaluation under six unseen image corruptions showed differences of at most 0.001 mIoU from the weather-oracle variant.