Fig 1.
The SeaDronesSee dataset contains real-world drone-captured images of maritime search and rescue scenarios, presenting unique challenges for object detection.
(a) Objects exhibit significant size and visual feature variations due to changes in drone flight altitude (Alt) and gimbal angle (GP), with substantial fluctuations in water surface patterns. (b) Close-ups of the scene show numerous small-sized targets with large size variations across different object categories, and overlap between boats, personnel, and equipment.
Table 1.
Sample counts and area statistics for object categories in the SeaDronesSee dataset.
Fig 2.
Overview of the proposed improvements to the YOLO11 model.
(a) SPD illustration; (b) YOLO11 backbone with four potential SPD integration points (1)-(4), where (1) and (2) were selected based on ablation experiments; (c) YOLO11 neck with three potential CARAFE replacement points [1]–[3], where [1] and [2] were selected based on ablation experiments; (d) YOLO11 head, with red dashed lines indicating the additional detection head for P2 features and the corresponding neck modules; (e) CARAFE illustration.
Table 2.
Ablation study results.
Table 3.
Comparison with YOLO-based state-of-the-art methods.
Fig 3.
Performance comparison between the proposed method and YOLO-based state-of-the-art methods on the SeaDronesSee dataset under varying model complexities.
(a) mAP curves for different parameter scales; (b) mAP curves for different computational complexities.
Table 4.
Comparison of the YOLO11 baseline and our optimized model at the “s” scale.
Fig 4.
Ground-truth annotations and detection results visualization of representative images from the SeaDronesSee dataset captured at different altitudes.