Object detection based on an adaptive attention mechanism.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Object detection is crucial in computer vision, with convolutional neural networks (CNNs) dominating recent advancements.
• CNN performance improvements have been achieved through architectural modifications like increased depth, width, cardinality, and skip/dense connections.
• Attention mechanisms, particularly channel-wise attention, have emerged as a significant area of research for enhancing CNNs.

Purpose of the Study:

• To investigate the benefits of combining global average pooling and global max pooling for channel-wise attention in CNNs.
• To propose novel adaptive attention units, including channel-wise, spatial-wise, and domain attention, to boost CNN performance.
• To introduce a new method for weighting attention vectors instead of simple concatenation.

Main Methods:

• Designed three novel attention units: adaptive channel-wise, adaptive spatial-wise, and adaptive domain attention.
• Developed a weighting strategy for attention vectors based on sub-unit outputs, differing from concatenation methods.
• Integrated the proposed attention mechanism into YOLOv3 and MobileNetv2 architectures.

Main Results:

• YOLOv3 with the proposed attention mechanism achieved performance gains of 2.9% and 1.2% mean average precision (mAP) on the KITTI and Pascal VOC datasets, respectively.
• MobileNetv2 integrated with the attention mechanism showed a 1.7% mAP improvement on the Pascal VOC dataset.
• The adaptive attention units demonstrated effectiveness in enhancing object detection capabilities of established CNN frameworks.

Conclusions:

• The proposed adaptive attention mechanisms offer a valuable enhancement for CNN-based object detection systems.
• Combining global average pooling and global max pooling within attention units can lead to improved feature representation.
• The developed attention strategy provides a more effective way to integrate multi-faceted attention information for better object detection results.