CrabNet: Fully Task-Specific Feature Learning for One-Stage Object Detection.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Object detection commonly uses shared deep learning backbones for classification and localization.
• Existing feature disentanglement methods often focus post-Region Proposal Network (RPN) and are limited to two-stage detectors.
• Limitations include insufficient task-specific feature learning before RPN and incompatibility with one-stage detectors.

Purpose of the Study:

• To propose a novel fully task-specific feature learning method for one-stage object detection.
• To overcome limitations of existing methods by enabling task-specific feature learning from the backbone stage.
• To enhance object detection performance in one-stage architectures.

Main Methods:

• Developed a method using two separated backbone models for disentangled classification and localization feature learning.
• Incorporated auxiliary heads at the end of each backbone to ensure fully task-specific features.
• Introduced a feature interaction module to align and fuse task-specific features for final detection.

Main Results:

• The proposed method, CrabNet, achieves clear performance improvements in object detection.
• CrabNet demonstrates favorable performance against state-of-the-art methods on the MS COCO dataset.
• The method shows increased efficiency with limited inference time.

Conclusions:

• Fully task-specific feature learning is effective for improving one-stage object detection.
• CrabNet offers a viable solution for enhancing detection accuracy and efficiency.
• The approach provides a new direction for feature disentanglement in deep learning-based object detection.