Deformable Capsules for Object Detection.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Capsule networks offer advantages over convolutional networks for representation but are computationally expensive and limited in modeling object pose.
• Existing capsule networks struggle with scalability for complex tasks like object detection due to memory and geometric constraints.

Purpose of the Study:

• To introduce a new family of capsule networks, deformable capsules (DeformCaps), specifically designed for object detection.
• To address the computational and memory limitations of previous capsule network architectures for large-scale vision tasks.

Main Methods:

• Proposed a novel capsule structure called SplitCaps.
• Introduced a new dynamic routing algorithm named SE-Routing.
• Developed a one-stage detection framework utilizing DeformCaps, SplitCaps, and SE-Routing.

Main Results:

• Achieved the first-ever capsule network for object detection.
• Demonstrated efficient scaling of capsule networks for large numbers of objects and classes.
• Obtained competitive results on the MS COCO dataset compared to state-of-the-art CNN-based methods.
• Showcased improved generalization to unusual object poses and viewpoints with fewer false positives.

Conclusions:

• DeformCaps, with SplitCaps and SE-Routing, provide an efficient and scalable capsule network solution for object detection.
• The proposed methods overcome previous limitations, enabling capsule networks to perform on par with CNNs in object detection tasks.
• This work opens new avenues for capsule networks in complex computer vision applications.