COME: A Collaborative Optimization Framework With Low-Rank MoE for Indoor 3D Object Detection.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Indoor 3D object detection is crucial for computer vision and robotics.
• Current methods often train domain-specific models, neglecting universal geometric attributes.
• This approach limits performance across diverse datasets.

Purpose of the Study:

• To propose COME, a collaborative optimization framework for indoor 3D object detection.
• To integrate universal geometric attributes while preserving domain-specific characteristics.
• To enhance cross-domain object detection performance.

Main Methods:

• COME utilizes a Cross-Domain Expert Parameter Sharing Strategy (CEPSS), inspired by Mixture of Experts (MoE).
• CEPSS features dual experts: domain-shared for universal attributes and domain-specific for unique features.
• A lightweight gating network dynamically selects experts, optimizing for different domains and reducing gradient conflicts. Low-rank structures enhance computational efficiency.

Main Results:

• COME achieves state-of-the-art results on benchmark datasets.
• The framework demonstrates superior performance compared to existing multi-domain detection methods.
• It shows acceptable parameter growth while improving detection accuracy.

Conclusions:

• COME effectively integrates universal and domain-specific features for enhanced 3D object detection.
• The proposed framework offers a significant advancement in cross-domain learning for computer vision tasks.
• COME provides a computationally efficient and high-performing solution for indoor 3D object detection.