Exploiting Gaussian based effective receptive fields for object detection.

Area of Science:

• Computer Vision
• Deep Learning
• Object Detection

Background:

• The effective receptive field (ERF) is vital for object detection, providing semantic information.
• Current methods use static ERF sizes, failing to account for image complexities like varying object scales.
• ERFs in real images often exhibit Gaussian distribution characteristics.

Purpose of the Study:

• To propose a dynamic, real-time, region-oriented ERF computation method.
• To enhance object detection by adapting ERF calculations to image content.
• To integrate this novel method into existing deep learning architectures.

Main Methods:

• Introduced Gaussian-based Effective Receptive Fields (GERF) for dynamic ERF computation.
• Applied GERF to the Bi-Level Routing Attention (BRA) module, creating GERF-BRA.
• Predicted ERFs per feature map window and weighted adjacent features using Gaussian distribution.
• Integrated GERF-BRA into YOLOv8n detection heads.

Main Results:

• GERF-BRA integration into YOLOv8n achieved a 2.5 AP improvement on the COCO 2017 dataset.
• Demonstrated significant effectiveness on proprietary agricultural and medical datasets.
• Validated the dynamic and region-oriented approach to ERF computation.

Conclusions:

• The proposed GERF method offers a dynamic and effective approach to ERF computation in object detection.
• GERF-BRA enhances object detection performance by better capturing object features.
• This method shows broad applicability across diverse datasets and domains.