Object detection in aerial images has always been a challenging task due to the generally small size of the objects. Most current detectors prioritize the development of new detection frameworks, often overlooking research on fundamental components such as feature pyramid networks. In this paper, we introduce the Cross-Layer Feature Pyramid Transformer (CFPT), a novel upsampler-free feature pyramid network designed specifically for small object detection in aerial images. CFPT incorporates two meticulously designed attention blocks with linear computational complexity: Cross-Layer Channel-Wise Attention (CCA) and Cross-Layer Spatial-Wise Attention (CSA). CCA achieves cross-layer interaction by dividing channel-wise token groups to perceive cross-layer global information along the spatial dimension, while CSA enables cross-layer interaction by dividing spatial-wise token groups to perceive cross-layer global information along the channel dimension. By integrating these modules, CFPT enables efficient cross-layer interaction in a single step, thereby avoiding the semantic gap and information loss associated with element-wise summation and layer-by-layer transmission. In addition, CFPT incorporates global contextual information, which improves detection performance for small objects. To further enhance location awareness during cross-layer interaction, we propose the Cross-Layer Consistent Relative Positional Encoding (CCPE) based on inter-layer mutual receptive fields. We evaluate the effectiveness of CFPT on three challenging object detection datasets in aerial images: VisDrone2019-DET, TinyPerson, and xView. Extensive experiments demonstrate that CFPT outperforms state-of-the-art feature pyramid networks while incurring lower computational costs. The code is available at https://github.com/duzw9311/CFPT.

As a crucial component of object detectors, current detection heads often lack the capability to effectively utilize contextual information, adapt to deformable objects, and align features and tasks. However, most existing methods prioritize a single capability, lacking comprehensive approaches to introduce them simultaneously. In this paper, we propose the Enhanced Head to integrate the above three capabilities into the detectors concurrently. Specifically, we propose three attention blocks with linear complexity: Global Concentrated Attention (GCA), Local Deformable Cross-Task Attention (LDCA), and Boundary-Aware Cross-Task Attention (BACA). The GCA captures long-range dependencies efficiently by employing Spatial Information Concentration (SIC). The LDCA improves feature alignment and deformation adaptability by enabling local deformable cross-task feature interactions. The BACA aligns classification features with localization results, enhancing task alignment and further improving deformation adaptability through a region-deformable interaction scheme. We implement Enhanced Head as a plug-and-play detection head and evaluate its effectiveness through extensive experiments on the MS COCO and VisDrone datasets. For instance, on the COCO detection benchmark, our Enhanced Head achieves +3.6 AP gain for FSAF, +3.3 AP for RetinaNet, and +2.9 AP for ATSS while reducing the FLOPs.

Feature upsampling is an essential operation in constructing deep convolutional neural networks. However, existing upsamplers either lack specific feature guidance or necessitate the utilization of high-resolution feature maps, resulting in a loss of performance and flexibility. In this paper, we find that the local self-attention naturally has the feature guidance capability, and its computational paradigm aligns closely with the essence of feature upsampling (i.e. feature reassembly of neighboring points). Therefore, we introduce local self-attention into the upsampling task and demonstrate that the majority of existing upsamplers can be regarded as special cases of upsamplers based on local self-attention. Considering the potential semantic gap between upsampled points and their neighboring points, we further introduce the deformation mechanism into the upsampler based on local self-attention, thereby proposing LDA-AQU. As a novel dynamic kernel-based upsampler, LDA-AQU utilizes the feature of queries to guide the model in adaptively adjusting the position and aggregation weight of neighboring points, thereby meeting the upsampling requirements across various complex scenarios. In addition, LDA-AQU is lightweight and can be easily integrated into various model architectures. We evaluate the effectiveness of LDA-AQU across four dense prediction tasks: object detection, instance segmentation, panoptic segmentation, and semantic segmentation. LDA-AQU consistently outperforms previous state-of-the-art upsamplers, achieving performance enhancements of 1.7 AP, 1.5 AP, 2.0 PQ, and 2.5 mIoU compared to the baseline models in the aforementioned four tasks, respectively.