Feature Correlation-Steered Capsule Network for object detection.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Convolutional Neural Networks (CNNs) excel at object detection but often neglect internal part-whole relationships.
• This oversight limits the understanding of feature correlations between objects and their parts, crucial for accurate detection.

Purpose of the Study:

• To propose a novel object detection approach that "looks inside" objects by leveraging part-whole feature correlations.
• To enhance object detection robustness by integrating Capsule Network (CapsNet) capabilities.

Main Methods:

• Introduction of Feature Correlation-Steered CapsNet (FCS-CapsNet).
• Utilizing Locally-Constrained Expectation-Maximum (EM) Routing Agreement (LCEMRA) for routing capsules.
• LCEMRA ensures only cohesive low-level capsules (parts) form high-level capsules (objects), uncovering part-object associations.

Main Results:

• FCS-CapsNet demonstrates promising object detection performance across multiple datasets (VOC2007, VOC2012, HKU-IS, DUTS, COCO).
• The method achieves results on par with state-of-the-art approaches.
• The proposed LCEMRA routing effectively captures part-object relationships.

Conclusions:

• The FCS-CapsNet effectively utilizes part-whole feature correlations for robust object detection.
• The novel LCEMRA routing mechanism improves the understanding of object composition.
• This approach offers a significant advancement in object detection by considering internal object structures.