Small-Object Sensitive Segmentation Using Across Feature Map Attention.

Area of Science:

• Computer Vision
• Deep Learning

Background:

• Semantic segmentation is crucial for scene understanding in autonomous driving.
• Deep Convolutional Neural Networks (CNNs) improve segmentation but struggle with small/thin objects due to information loss from convolutional and pooling operations.

Purpose of the Study:

• To address the challenge of segmenting small/thin objects in semantic segmentation.
• To introduce a novel attention-based method, Across Feature Map Attention (AFMA), to enhance segmentation accuracy for these challenging objects.

Main Methods:

• Developed Across Feature Map Attention (AFMA), an attention-based technique.
• AFMA quantifies relationships between small and large objects across different feature levels.
• The method acts as a plug-in for existing segmentation architectures.

Main Results:

• AFMA compensates for high-level feature loss in small objects.
• Substantially and consistently improves segmentation of small/thin objects.
• Provides more interpretable feature representations compared to previous methods.

Conclusions:

• AFMA is an effective method for improving small/thin object segmentation.
• The approach enhances existing semantic segmentation architectures.
• Demonstrated significant improvements on CamVid and Cityscapes datasets.