MSFFAL: Few-Shot Object Detection via Multi-Scale Feature Fusion and Attentive Learning

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Few-shot object detection (FSOD) addresses limitations of traditional detectors in low-data scenarios.
• Meta-learning methods show promise but struggle with active identification and small object detection.
• Current methods face challenges with feature selection and representation of hard samples.

Purpose of the Study:

• To introduce a novel framework, Multi-scale Feature Fusion and Attentive Learning (MSFFAL), for few-shot object detection.
• To enhance the model's ability for active identification and improve small object detection accuracy.
• To optimize the testing process by removing the dependency on query labels for feature selection.

Main Methods:

• Designed a backbone incorporating multi-scale feature fusion and a channel attention mechanism.
• Developed an attention loss function to replace traditional feature weighting modules.
• Enabled consistent representation of objects across different branches for improved feature learning.

Main Results:

• MSFFAL achieves state-of-the-art (SOTA) performance, outperforming existing methods by 0.7-7.8% on Pascal VOC.
• Demonstrated a 1.61x improvement in detecting small objects on the MS COCO dataset compared to the baseline.
• The proposed attention loss facilitates active recognition and optimizes the testing phase.

Conclusions:

• The MSFFAL framework effectively improves few-shot object detection accuracy, particularly for small objects.
• The attention loss mechanism enables label-free feature selection during testing, enhancing practical applicability.
• MSFFAL represents a significant advancement in addressing key challenges within few-shot object detection research.