Object detectors involving a NAS-gate convolutional module and capsule attention module.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• State-of-the-art object detectors rely on backbone architecture modifications and feature pyramids for performance.
• Optimizing feature representation is crucial for enhancing object detection capabilities.

Purpose of the Study:

• To explore modifications of object detector backbones and feature pyramids using Neural Architecture Search (NAS) and Capsule Networks.
• To introduce and evaluate novel modules for improved object detection performance.

Main Methods:

• Developed a NAS-gate convolutional module for backbone optimization, addressing object scale variation.
• Introduced a Capsule Attention module to generate spatial attention masks for feature pyramid enhancement.
• Integrated these modules into NASGC-CapANet, a novel object detection framework.

Main Results:

• The NAS-gate convolutional module effectively alleviates object scale variation issues.
• The Capsule Attention module improves feature representation and detector localization accuracy.
• NASGC-CapANet-based Faster R-CNN achieved 2.7% and 2.0% higher mAPs than baseline ResNet-50 and ResNet-101 backbones, respectively.
• NASGC-CapANet-based Cascade R-CNN reached a 43.8% box mAP on the MS COCO test-dev dataset.

Conclusions:

• The proposed NASGC-CapANet framework significantly enhances object detection performance.
• The integration of NAS and Capsule Networks offers a promising direction for future object detection research.
• The developed modules provide effective solutions for object scale variation and localization inaccuracies.