Context-Aware Enhanced Feature Refinement for small object detection with Deformable DETR.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Small object detection is crucial for autonomous driving and surveillance.
• Existing Deformable DETR models struggle with small objects due to CNN limitations in global context and feature representation.
• Significant size disparities in datasets hinder the detection of objects with few pixels.

Purpose of the Study:

• To enhance the Deformable DETR network for improved small object detection.
• To address limitations in feature extraction and small object representation.
• To improve performance in critical applications like autonomous driving.

Main Methods:

• Proposed a Context-Aware Enhanced Feature Refinement Deformable DETR.
• Introduced Mask Attention in the backbone for better feature extraction and background suppression.
• Developed a Context-Aware Enhanced Feature Refinement Encoder to handle limited pixel representation of small objects.

Main Results:

• The proposed method significantly outperforms the baseline Deformable DETR.
• Achieved a 2.1% improvement in mean Average Precision (mAP).
• Demonstrated enhanced capability in detecting small objects.

Conclusions:

• The Context-Aware Enhanced Feature Refinement Deformable DETR effectively improves small object detection.
• Mask Attention and the novel encoder contribute to superior feature representation and detection accuracy.
• The approach offers a promising solution for applications requiring robust small object detection.