Expandable-RCNN: toward high-efficiency incremental few-shot object detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Incremental Few-Shot Object Detection (iFSOD) presents challenges in sequentially learning new object categories with limited data.
• Existing methods struggle with overfitting and category bias in adaptive detection systems.

Purpose of the Study:

• To propose an efficient and simple framework, Expandable-RCNN, for addressing the iFSOD problem.
• To enable online sequential addition of new classes without retraining the base network.

Main Methods:

• Adapted Faster R-CNN with two novel components: an IOU-aware weight imprinting strategy and a group soft-max layer (GSL) for bias correction.
• The IOU-aware imprinting avoids overfitting by directly determining classifier weights for new classes and background.
• The GSL module calibrates biased predictions to improve classification performance and prevent catastrophic forgetting.

Main Results:

• Expandable-RCNN demonstrated significant improvements on the MS-COCO dataset.
• The proposed method outperformed the state-of-the-art method ONCE by 5.9 points on few-shot classes.
• Achieved effective online adaptive detection with zero retraining of the base network.

Conclusions:

• Expandable-RCNN offers an effective solution for the iFSOD problem, enabling efficient online learning of new object categories.
• The combination of IOU-aware weight imprinting and GSL successfully mitigates overfitting and category bias.
• The framework provides a robust approach for adaptive detection systems requiring sequential learning capabilities.