Improved region proposal network for enhanced few-shot object detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning object detection requires extensive annotated data, which is costly and time-consuming, especially for rare objects.
• Few-shot object detection (FSOD) reduces data needs but struggles with novel class instances appearing as background noise, degrading performance.

Purpose of the Study:

• To develop a semi-supervised algorithm to detect and leverage unlabeled novel objects in FSOD training.
• To improve FSOD performance by mitigating the negative impact of novel class instances on base models.

Main Methods:

• A hierarchical ternary classification region proposal network (HTRPN) was developed to localize and label unlabeled novel objects.
• An improved hierarchical sampling strategy for the region proposal network (RPN) was implemented to enhance the detection of large objects.

Main Results:

• The proposed method effectively detects and utilizes unlabeled novel objects as positive samples during FSOD training.
• Experimental results on COCO and PASCAL VOC datasets demonstrate superior performance compared to existing state-of-the-art FSOD methods.

Conclusions:

• The developed semi-supervised approach significantly enhances few-shot object detection capabilities.
• The HTRPN method offers a promising solution for improving object detection with limited annotated data and handling novel object classes.