Dynamic Graph Message Passing Networks.

Area of Science:

• Computer Vision
• Deep Learning
• Graph Neural Networks

Background:

• Convolutional Neural Networks (CNNs) struggle with long-range dependencies in scene understanding.
• Fully-connected graphs, like those in Transformers, are effective but computationally expensive.

Purpose of the Study:

• To propose a dynamic graph message passing network that efficiently models long-range dependencies.
• To reduce computational complexity compared to fully-connected graph approaches.
• To develop a Transformer-based backbone for various computer vision tasks.

Main Methods:

• Developed a dynamic graph message passing network with adaptive node sampling.
• Dynamically predicted node-dependent filter weights and affinity matrices.
• Designed a self-attention module and a Transformer-based backbone network.

Main Results:

• Achieved significant improvements on image classification, object detection, and segmentation tasks.
• Outperformed state-of-the-art baselines and fully-connected graphs.
• Demonstrated substantial reductions in floating-point operations and parameters.

Conclusions:

• The proposed dynamic graph network offers an efficient and effective solution for modeling long-range dependencies in computer vision.
• This approach provides a strong alternative to computationally intensive fully-connected graph methods.
• The developed Transformer-based backbone shows versatility across multiple downstream vision tasks.