P2LITE-YOLO: DETECTION-SCALE REALLOCATION FOR TINY OBJECT DETECTION IN UAV IMAGERY
Keywords:
UAV imagery, Tiny object detection, Scale prior, Resolution-semantic collaboration, Prediction-capacity reallocationAbstract
In UAV imagery, object scales are heavily biased towards the tiny region, yet general-purpose object detectors typically employ fixed multi-scale prediction configurations. There is a significant mismatch between the spatial detail loss caused by continuous downsampling and the prediction scale setup. To address this issue, this paper proposes P2Lite-YOLO, a scale-prior-driven prediction resource reallocation method for tiny UAV objects. First, the prediction requirements of different resolution features are explicitly characterized based on the object scale distribution in the training data. While keeping the number of detection heads unchanged, the original P3/P4/P5 prediction scales of YOLO11s are reallocated to P2/P3/P4, allowing the limited prediction capacity to prioritize high-resolution regions where tiny objects dominate. Second, a P2 collaborative representation is constructed via top-down deep semantic propagation and shallow spatial detail fusion, enabling high-resolution features to retain both fine-grained localization information and contextual discriminative ability. Simultaneously, a detection head capacity alignment strategy is adopted to mitigate the impact of structural adjustments on pre-trained parameter transfer. Statistics from the VisDrone2019 training set show that 68.30% of the 343,205 valid objects have an equivalent size of less than 16 pixels, confirming the necessity of allocating high-resolution prediction resources. Under a unified experimental protocol, P2Lite-YOLO achieves 43.11% mAP50 and 26.38% mAP50-95 with 9.129 M parameters, outperforming YOLO11s by 2.88 and 2.47 percentage points, respectively, and surpassing comparative models such as YOLO11m, YOLOv8m, Faster R-CNN, and RetinaNet. Further experiments indicate that retaining the independent P5 prediction branch after introducing the P2 high-resolution prediction does not yield performance gains. These results demonstrate that, compared to simply expanding the detection scales, reconfiguring multi-scale representation and prediction resources under a fixed prediction budget based on data scale priors is an effective strategy for improving tiny object detection in UAV imagery.References
[1] Zhu P, Wen L, Du D, et al. Detection and tracking meet drones challenge. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 44(11): 7380-7399. DOI: 10.1109/TPAMI.2021.3119563.
[2] Tang G, Ni J, Zhao Y, et al. A survey of object detection for UAVs based on deep learning. Remote Sensing, 2024, 16(1): 149. DOI: 10.3390/rs16010149.
[3] Ren S, He K, Girshick R, et al. Faster R-CNN: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6): 1137-1149. DOI: 10.1109/TPAMI.2016.2577031.
[4] Wang C Y, Liao H Y M. YOLOv1 to YOLOv10: The fastest and most accurate real-time object detection systems. APSIPA Transactions on Signal and Information Processing, 2024, 13(1): 1-38. DOI: 10.1561/116.20240058.
[5] Lin T Y, Dollár P, Girshick R, et al. Feature pyramid networks for object detection. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017: 2117-2125.
[6] Liu S, Qi L, Qin H, et al. Path aggregation network for instance segmentation. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018: 8759-8768.
[7] Wang C Y, Liao H Y M, Wu Y H, et al. CSPNet: A new backbone that can enhance learning capability of CNN. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2020: 390-391.
[8] Ge Z, Liu S, Wang F, et al. YOLOX: Exceeding YOLO series in 2021. arXiv:2107.08430, 2021.
[9] Terven J, Córdova-Esparza D M, Romero-González J A. A comprehensive review of YOLO architectures in computer vision: From YOLOv1 to YOLOv8 and YOLO-NAS. Machine Learning and Knowledge Extraction, 2023, 5(4): 1680-1716. DOI: 10.3390/make5040083.
[10] Khanam R, Hussain M. YOLOv11: An overview of the key architectural enhancements. arXiv:2410.17725, 2024.
[11] Zhu X, Lyu S, Wang X, et al. TPH-YOLOv5: Improved YOLOv5 based on transformer prediction head for object detection on drone-captured scenarios. Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, 2021: 2778-2788.
[12] Xu C, Wang J, Yang W, et al. Detecting tiny objects in aerial images: A normalized Wasserstein distance and a new benchmark. ISPRS Journal of Photogrammetry and Remote Sensing, 2022, 190: 79-93. DOI: 10.1016/j.isprsjprs.2022.06.002.
[13] Zhao J D, Zhen G Y, Chu C Q. Unmanned aerial vehicle image target detection algorithm based on YOLOv8. Computer Engineering, 2024, 50(4): 113-120. DOI: 10.19678/j.issn.1000-3428.0068268.
[14] Li X, Wang W, Wu L, et al. Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. Advances in Neural Information Processing Systems, 2020, 33: 21002-21012.
[15] Lin T Y, Goyal P, Girshick R, et al. Focal loss for dense object detection. Proceedings of the IEEE International Conference on Computer Vision, 2017: 2980-2988.