Object detection is challenging in high spatial resolution (HSR) remote sensing images that have a complex background and irregular object locations. To minimize manual annotation cost in supervised learning methods and achieve advanced detection performance, we proposed a point-based weakly supervised learning method to address the object detection challenge in HSR remote sensing images. In the study, point labels are introduced to guide candidate bounding box mining and generate pseudobounding boxes for objects. Then, pseudobounding boxes are applied to train the detection model. A progressive candidate bounding box mining strategy is proposed to refine object detection. Experiments are conducted on a comprehensive HSR dataset which contains four categories. Results indicate the proposed method achieves competitive performance compared to YOLOv5 which is trained on manual bounding box annotations. In comparison to the state-of-the-art weakly supervised learning method, our method outperforms WSDDN method with 0.62 mean average precision score.
Point-Based Weakly Supervised Learning for Object Detection in High Spatial Resolution Remote Sensing Images / Li, Y.; He, B.; Melgani, F.; Long, T.. - In: IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING. - ISSN 1939-1404. - 14:(2021), pp. 5361-5371. [10.1109/JSTARS.2021.3076072]
Point-Based Weakly Supervised Learning for Object Detection in High Spatial Resolution Remote Sensing Images
Melgani F.;
2021-01-01
Abstract
Object detection is challenging in high spatial resolution (HSR) remote sensing images that have a complex background and irregular object locations. To minimize manual annotation cost in supervised learning methods and achieve advanced detection performance, we proposed a point-based weakly supervised learning method to address the object detection challenge in HSR remote sensing images. In the study, point labels are introduced to guide candidate bounding box mining and generate pseudobounding boxes for objects. Then, pseudobounding boxes are applied to train the detection model. A progressive candidate bounding box mining strategy is proposed to refine object detection. Experiments are conducted on a comprehensive HSR dataset which contains four categories. Results indicate the proposed method achieves competitive performance compared to YOLOv5 which is trained on manual bounding box annotations. In comparison to the state-of-the-art weakly supervised learning method, our method outperforms WSDDN method with 0.62 mean average precision score.File | Dimensione | Formato | |
---|---|---|---|
JSTARS-2021-Object Detection.pdf
accesso aperto
Tipologia:
Versione editoriale (Publisher’s layout)
Licenza:
Creative commons
Dimensione
7.58 MB
Formato
Adobe PDF
|
7.58 MB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione