Context-Aware Block Net for Small Object Detection.

Area of Science:

• Computer Vision
• Deep Learning
• Object Detection

Background:

• Object detectors often lose spatial information due to downsampling, hindering small object detection.
• Existing methods struggle with representing small objects effectively in downsampled feature maps.

Purpose of the Study:

• To propose a novel Context-Aware Block Net (CAB Net) for improved small object detection.
• To enhance the representation capacity of high-resolution feature maps for better semantic understanding.

Main Methods:

• Designed a Context-Aware Block (CAB) utilizing pyramidal dilated convolutions.
• Integrated CAB into a truncated backbone network with a small downsampling factor.
• Developed CAB Net to generate high-resolution, semantically rich feature maps.

Main Results:

• CAB Net demonstrated superior performance on the Tsinghua-Tencent 100K and Airport datasets.
• The proposed method significantly outperformed existing top-performing object detectors.
• Achieved real-time detection speeds, validating the efficiency of CAB Net.

Conclusions:

• CAB Net effectively improves small object detection by preserving spatial and semantic information.
• The architecture successfully captures both low-level patterns and high-level semantics of small objects.
• CAB Net offers a promising solution for real-time, high-accuracy small object detection.