Qiang Fu

Class imbalance is a fundamental challenge in semantic segmentation of remote sensing imagery, causing deep convolutional networks to systematically neglect minority categories such as vehicles and small water bodies. To address this, we propose a class distribution-aware adversarial training framework that explicitly penalizes the model’s bias toward dominant classes at the representation level. The framework comprises two key components. First, a Hierarchical Attention U-Net (HAU-Net) integrates a Parallel Hybrid Attention Module (PHAM) at multiple scales, combining spatial and channel attention to amplify feature responses for small and sparsely distributed objects. Second, a distribution discriminator is trained to distinguish the predicted per-class probability distribution from a balanced uniform prior; through a minimax game, the segmentation network is forced to produce class-balanced predictions without manual loss re-weighting. Extensive experiments on the ISPRS Potsdam and Vaihingen datasets demonstrate that our method achieves an overall mIoU of 73.82% and delivers substantial improvements on severely under-represented minority classes, including a +12.11% IoU gain on vehicles and a +19.26% IoU gain on clutter/background, compared to our non-adversarial baseline. Ablation studies confirm that both the hierarchical attention design and the adversarial training contribute independently to these improvements. Our framework offers an effective solution to class imbalance in remote sensing segmentation, with the potential to be integrated into diverse segmentation architectures.

Camouflaged person detection from unmanned aerial vehicle (UAV) views is important for intelligent surveillance, search and rescue, and battlefield perception, yet remains insufficiently studied. This task is particularly challenging due to small target scales, cluttered backgrounds, and large viewpoint variations in UAV imagery. Existing public datasets mainly focus on conventional pedestrian scenarios or generic camouflaged objects, and therefore cannot adequately support benchmarking for UAV-view camouflaged person detection. To address this gap, we present UAV-CPD, a benchmark dataset dedicated to camouflaged person detection from UAV views, and establish a unified evaluation protocol for this task. We further propose CamouDETR, a frequency-aware end-to-end baseline built upon Dual-Branch Wavelet Enhancement (DWE), Dynamic Frequency Routing (DFR), and the Context-Aware Reference Decoder (CARD), to strengthen weak-detail representation, cross-scale frequency transmission, and context-aware box refinement. Experiments show that CamouDETR achieves 77.30 mAP on UAV-CPD, outperforming RT-DETR by 3.10 points, and also improves mAP on AMAD-5 from 62.16 to 65.40. These results support UAV-CPD as a useful benchmark for this task and show that the proposed baseline offers a practical starting point for UAV-view camouflaged person detection.

Reliable 3D multi-object tracking (MOT) is essential for autonomous driving. The model-based Kalman filter (KF) is the standard motion estimator in 3D MOT, but its linear-dynamics assumption and constant noise covariances limit its adaptability to nonlinear motion and time-varying noise. The learning-aided Kalman filter (LAKF) addresses these issues by learning the Kalman gain (KG) through recurrent neural networks. We adopt CNPE-KalmanNet, a variant that integrates complex-number position encoding to capture semantic correlations among heterogeneous state components. However, the analytical KG is derived under the assumption that the process and measurement noise covariance matrices are positive definite, whereas the learned KG from CNPE-KalmanNet lacks structural constraints. The unconstrained learned KG tends to produce a dense matrix, leading to cross-dimensional coupling, particularly across different scales. This paper proposes a hybrid KG that fuses the model-based analytical KG with the data-driven learned KG. Compared to the AB3DMOT baseline on KITTI and the P3DMOT baseline on nuScenes, our approach improves average recall by 13% and 22%, and IoU by 6% and 7% in state estimation, and AMOTA by 0.7% and 0.5% for overall tracking accuracy. The hybrid design retains the numerical stability of analytical filtering while incorporating the strong nonlinear fitting capacity from deep learning.

Object detection from aerial platforms under adverse atmospheric conditions, particularly haze, is paramount to robust drone autonomy. Yet, this domain remains largely underexplored, primarily hindered by the absence of specialized benchmarks. To bridge this gap, we present HazyDet, the first, large-scale benchmark specifically designed for drone-view object detection in hazy conditions. Comprising 383,000 real-world instances derived from both naturally hazy captures and synthetically hazed scenes augmented from clear images, HazyDet provides a challenging and realistic testbed for advancing detection algorithms. To address the severe visual degradation induced by haze, we propose the Depth-Conditioned Detector (DeCoDet), a novel architecture that integrates a Depth-Conditioned Kernel to dynamically modulate feature representations based on depth cues. The practical efficacy and robustness of DeCoDet are further enhanced by its training with a Progressive Domain Fine-Tuning (PDFT) strategy to mitigate synthetic-to-real domain shifts, and a Scale-Invariant Refurbishment Loss (SIRLoss) to ensure resilient learning from potentially noisy depth annotations. Comprehensive empirical validation on HazyDet substantiates the superiority of our unified DeCoDet framework, which achieves state-of-the-art performance, surpassing the closest competitor by a notable +1.0% mAP on challenging real-world hazy test scenarios. Our dataset and toolkit are available at https://github.com/GrokCV/HazyDet.

Reliable detection of unmanned aerial vehicle (UAV) swarms is essential for airspace security and defense applications, yet the scarcity of large-scale, densely annotated training data remains a critical bottleneck. Collecting real-world swarm data is costly, logistically demanding, and constrained by airspace regulations, while manual annotation of numerous small, fast-moving targets is time-consuming and error-prone. To address these challenges, this paper presents SynthSwarm, a large-scale synthetic dataset specifically designed for UAV swarm detection in long-range aerial surveillance scenarios. The dataset is generated through a controllable simulation pipeline built on the Unity engine, which enables precise six-degree-of-freedom (6-DoF) pose specification for each UAV instance and automatic pixel-accurate bounding box annotation without manual labeling. SynthSwarm comprises 7,000 high-resolution images (1920$\times$1080) containing 31,542 UAV instances, with systematic variations in swarm density, formation patterns, target scale, and environmental conditions. Statistical analysis reveals that 67.3\% of the targets qualify as small objects, reflecting the inherent difficulty of detecting distant UAV swarms. We benchmark several representative deep learning detectors, including YOLOX,YOLOv13,YOLOv12,YOLOv6 on the proposed dataset. Experimental results demonstrate that the dataset poses significant challenges for existing methods, particularly in high-density and small-target scenarios. Furthermore, cross-dataset experiments validate the effectiveness of synthetic data as a pre-training source for improving detection performance on real UAV datasets. The dataset and generation pipeline are publicly available to facilitate further research in UAV swarm detection.

KalmanNet demonstrates potential for nonlinear state estimation, but it directly concatenates heterogeneous components with disparate semantics into a composite vector. This approach ignores capturing semantic component correlations, limiting the network's ability to utilize this information when computing the Kalman gain. To address this, we propose CNPE-KalmanNet, a novel learning-aided Kalman filter that models semantic component correlations by integrating element-wise rotary encoding with multilayer perceptrons in the complex domain. CNPE serves as a plug-and-play module compatible with two baseline KalmanNet architectures. Experiments on the Michigan NCLT dataset demonstrate that CNPE-KalmanNet significantly improves estimation accuracy. Compared to the existing KalmanNet, our method reduces the RMSE by 20%-25% in state estimation tasks and by 15%-20% in sensor fusion tasks.

Motion estimation is a crucial component in multi-object tracking (MOT). It predicts the trajectory of objects by analyzing the changes in their positions in consecutive frames of images, reducing tracking failures and identity switches. The Kalman filter (KF) based on the linear constant-velocity model is one of the most commonly used methods in MOT. However, it may yield unsatisfactory results when KF's parameters are mismatched and objects move in non-stationary. In this work, we utilize the learning-aided filter to handle the motion estimation of MOT. In particular, we propose a novel method named Semantic-Independent KalmanNet (SIKNet), which encodes the state vector (the input feature) using a Semantic-Independent Encoder (SIE) by two steps. First, the SIE uses a 1D convolution with a kernel size of 1, which convolves along the dimension of homogeneous-semantic elements across different state vectors to encode independent semantic information. Then it employs a fully-connected layer and a nonlinear activation layer to encode nonlinear and cross-dependency information between heterogeneous-semantic elements. To independently evaluate the performance of the motion estimation module in MOT, we constructed a large-scale semi-simulated dataset from several open-source MOT datasets. Experimental results demonstrate that the proposed SIKNet outperforms the traditional KF and achieves superior robustness and accuracy than existing learning-aided filters. The code is available at (https://github.com/SongJgit/filternet and https://github.com/SongJgit/TBDTracker).

Identifying challenging samples in real-time aerial images captured by UAVs is currently a difficult task. Lightweight network designs often utilize lightweight modules or employ methods such as network pruning, quantization, and distillation for real-time detection on embedded devices. However, these approaches are inadequate for complex background aerial images. In this paper, we propose a real-time object detection algorithm specifically designed for complex aerial images, named MINIAOD. First, we use Ghost convolution to build a lightweight backbone network, which improves the detection speed without compromising the feature extraction capability. Additionally, we utilize the GSConv module to construct a feature enhancement network and develop the GSC3ECA module to improve the network’s learning capability for challenging targets, while also decreasing computational complexity. To boost the network’s focus on small and medium-sized targets and mitigate interference from complex backgrounds, we incorporate the ECA attention mechanism in the small and medium-sized target detection branch to elevate the focus on difficult-to-detect samples. Finally, we design the rotation detection head to characterize the target directionality of aerial images. We perform extensive experiments on multiple publicly available datasets and the proposed method demonstrates a favorable balance between accuracy and speed when compared to the current leading models.