Towards Stabilized Few-Shot Object Detection with Less Forgetting via Sample Normalization.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Few-shot object detection (FSOD) aims to identify novel object classes with limited labeled data.
• Existing FSOD methods often suffer from performance instability and forgetting during meta-training.
• Gaps in meta-knowledge transfer contribute to these limitations in meta-learning frameworks.

Purpose of the Study:

• To address the challenges of forgetting and performance instability in few-shot object detection.
• To enhance meta-knowledge transfer for more robust few-shot learning.
• To improve the stability and accuracy of object detection models trained with limited data.

Main Methods:

• Proposed a novel method called sample normalization to improve performance stability and decrease forgetting.
• Applied Z-score normalization to mitigate the hubness problem in high-dimensional feature spaces.
• Evaluated the approach on the PASCAL VOC dataset for few-shot object detection tasks.

Main Results:

• The proposed sample normalization method significantly outperforms existing methods in accuracy and stability.
• Achieved substantial improvements in mean Average Precision (mAP@0.5) and mean Recall (mAR) in single runs and across multiple experiments.
• Demonstrated alleviation of performance degradation on base classes, indicating better generalization.

Conclusions:

• Sample normalization is an effective technique for enhancing stability and reducing forgetting in few-shot object detection.
• The method improves meta-knowledge transfer, leading to superior performance compared to current state-of-the-art approaches.
• The findings suggest a promising direction for developing more robust and reliable few-shot learning systems.