Sampling Equivariant Self-Attention Networks for Object Detection in Aerial Images

Area of Science:

• Computer Vision
• Deep Learning
• Image Analysis

Background:

• Aerial object detection is challenging due to significant variations in object scale and orientation.
• Standard deep convolutional neural networks struggle with these variations.
• Existing methods like deformable convolutional networks offer limited sampling equivariance.

Purpose of the Study:

• To develop a novel deep learning model for improved aerial object detection.
• To enhance sampling equivariance in neural networks for robust feature extraction.
• To achieve state-of-the-art performance on aerial image datasets.

Main Methods:

• Proposed sampling equivariant self-attention networks using local image patches and masks for sampling.
• Introduced a transformation embedding module to further improve equivariant sampling.
• Developed a randomized normalization module for enhanced network generalization.
• Created a quantitative evaluation metric for sampling equivariance.

Main Results:

• The proposed model demonstrates significantly superior sampling equivariance compared to existing methods.
• Achieved state-of-the-art results on DOTA-v1.0, DOTA-v1.5, and HRSC2016 datasets.
• The model extracts more effective image features without additional computation or parameters.

Conclusions:

• Sampling equivariant self-attention networks offer a robust solution for aerial object detection.
• The proposed methods enhance network generalization and provide fair evaluation of sampling equivariance.
• This approach advances the field of object detection in challenging aerial imagery.