End-to-end DEtection TRansformes (DETRs) are leading a new trend in various object detection tasks. However, when it comes to the ubiquitous tiny objects in aerial imagery, the potential of DETRs still remains under-explored. In this work, we observe that the expansive field of view of remote sensing images often results in a limited pixel representation of tiny objects coupled with a substantial variance in the number of instances across images. The significantly varied tiny object number per image conflicts with DETRs' fixed set of object queries. A large number of queries are necessary to ensure high recall in dense scenarios, while sparse scenarios benefit from fewer, more distinct queries. To tackle this issue, we propose a Density-aware DETR with Dynamic Query (D3Q). D3Q adaptively determines the optimal number of object queries for each image by explicitly estimating its object density. This dynamic query mechanism enables efficient and accurate tiny object detection under both dense and sparse object distributions. Additionally, we introduce a refined box loss designed for tiny object detection that further stabilizes training. Through these strategies, D3Q effectively adapts to both dense and sparse scenarios, overcoming the limitations of fixed query in DETR. Extensive experiments on challenging tiny object detection benchmarks demonstrate the superior performance of D3Q compared to state-of-the-art methods. Particularly, when integrated with DINO, D3Q achieves an impressive 32.1% mAP on the AI-TOD-v2 dataset, setting a new state-of-the-art performance.

Detecting oriented tiny objects, which are limited in appearance information yet prevalent in real-world applications, remains an intricate and under-explored problem. To address this, we systematically introduce a new dataset, a benchmark, and a dynamic coarse-to-fine learning scheme in this study. Our proposed dataset, AI-TOD-R, features the smallest object sizes among all oriented object detection datasets. Based on AI-TOD-R, we present a benchmark spanning a broad range of detection paradigms, including both fully-supervised and label-efficient approaches. Through investigation, we identify a learning bias presents across various learning pipelines: confident objects become increasingly confident, while vulnerable oriented tiny objects are further marginalized, hindering their detection performance. To mitigate this issue, we propose a Dynamic Coarse-to-Fine Learning (DCFL) scheme towards unbiased learning. DCFL dynamically updates prior positions to better align with the limited areas of oriented tiny objects, and it assigns samples in a way that balances both quantity and quality across different object shapes, thus mitigating biases in prior settings and sample selection. Extensive experiments across 10 challenging object detection datasets demonstrate that DCFL achieves state-of-the-art accuracy, high efficiency, and remarkable versatility.

Remote sensing (RS) imagery captures the earth's ever-changing landscapes, reflecting evolving land cover patterns propelled by natural processes and human activities. However, existing RS scene classification methods mainly operate under a closed-set hypothesis, which stumbles when encountering novel emerging scenes. This article addresses the intricate task of RS scene classification without labels for novel scenes under incremental learning, termed class-incremental novel category discovery. We propose a contrastive learning-based novel category discovery pipeline tailored for RS image scene classification, enhancing the ability to learn unlabeled novel class data. Furthermore, within this pipeline, we introduce a positive pair filter to identify more positive sample pairs from novel classes, improving the feature representation capability on unlabeled data. Besides, our contrastive learning pipeline incorporates an old-feature replaying method to alleviate catastrophic forgetting in old classes. Extensive evaluations across three public RS datasets showcase the superiority of our method over state-of-the-art approaches.

Aerial images record the dynamic Earth terrain, reflecting changes in land cover patterns caused by natural processes and human activities. Nonetheless, prevailing aerial image classification methodologies predominantly function within a closed-set framework, thereby encountering challenges when confronted with the identification of newly emerging scenes. To address this, this paper explores an aerial image recognition scenario in which a dataset comprises both labeled and unlabeled aerial images, intending to classify all images within the unlabeled subset, termed Generalized Category Discovery (GCD). It is noteworthy that the unlabeled images may pertain to labeled classes or represent novel classes. Specifically, we first develop a contrastive learning framework drawing upon the cutting-edge algorithms in GCD. Based on the multi-object characteristics of aerial images, we then propose a slot attention-based GCD training process (Slot-GCD) that contrasts learning at both the object and image levels. It decouples multiple local object features from feature maps using slots and then reconstructs the overall semantic feature of the image based on slot confidence scores and the feature map. Finally, these object-level and image-level features are input into the contrastive learning module to enable the model to learn more precise image semantic features. Comprehensive evaluations across three public aerial image datasets highlight the superiority of our approach over state-of-the-art methods. Particularly, Slot-GCD achieves a recognition accuracy of 91.5% for known old classes and 81.9% for unknown novel class data on the AID dataset.

Objects in aerial images tend to be densely scattered and appear in arbitrary orientations, making the annotation process quite costly. To reduce the annotation cost, existing methods propose randomly annotating a proportion of images or objects for aerial object detection with fewer label usage. These approaches, however, can lead to redundancy in labels and inherit the biases associated with the imbalance in datasets. To minimize sample redundancy and alleviate data imbalance, we propose a novel labeling pattern that acquires heterogeneous object labels in a class-orthogonal manner, preserving a broader diversity of samples for each category with less annotation effort. To improve data utility, we design a Dynamic Multi-View Learning (DML) strategy to overcome the sample quantity-quality dilemma in current pseudo-labeling methods—a high pseudo-label threshold reduces sample quantity, while low thresholds compromise sample quality. First, DML separates model predictions into multiple hierarchies for finer screening, mitigating the suppression of unlabeled objects in binary pseudo-label strategies. With this separation, DML learns to construct a new view by injecting high-quality samples and masking low-quality regions in this view, simultaneously expanding sample quantity while ensuring sample quality. Unlike previous methods that mine pseudo labels solely from unlabelled regions, DML releases this constraint by learning to expand high-quality samples with a dynamic view. Extensive experiments on five benchmark datasets validate our method’s state-of-the-art accuracy and label efficiency. Notably, with approximately 5% DOTA-v2.0 annotations, DML achieves nearly 90% of the fully supervised performance. The codes will be available at https://github.com/ZhangRuixiang-WHU/ALOD_DML/.