CBNet: A Composite Backbone Network Architecture for Object Detection.

Area of Science:

• Computer Vision
• Deep Learning
• Object Detection

Background:

• High-performing object detectors rely heavily on advanced backbone networks.
• Existing backbone architectures often require extensive modifications for optimal performance.
• There is a need for flexible and efficient backbone frameworks that leverage pre-trained models.

Purpose of the Study:

• To propose a novel and flexible backbone framework, CBNet, for constructing high-performance object detectors.
• To enhance object detection by integrating multi-level features and expanding receptive fields.
• To demonstrate the generalization capabilities of CBNet across various backbone and detector designs.

Main Methods:

• Developed CBNet, a framework that groups multiple identical backbones connected through composite connections.
• Integrated high- and low-level features from multiple backbone networks.
• Implemented a training strategy with auxiliary supervision for CBNet-based detectors.

Main Results:

• CB-Swin-L achieved state-of-the-art results on COCO test-dev: 59.4% box AP and 51.6% mask AP (single-scale).
• Reduced training time by 6x compared to the baseline Swin-L.
• Achieved new record single-model results of 60.1% box AP and 52.3% mask AP with multi-scale testing.

Conclusions:

• CBNet offers a more efficient, effective, and resource-friendly method for building high-performance backbone networks.
• The framework demonstrates strong generalization across different backbone types (CNN, Transformer) and detector designs (one-stage, two-stage).
• CBNet enables significant performance gains without requiring additional pre-training of the composite backbone.