Learning Representative Features by Deep Attention Network for 3D Point Cloud Registration.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Three-dimensional point cloud registration is crucial for aligning spatial data in applications like underground mining.
• Learning-based methods, especially attention models, show promise but face computational challenges.
• Existing encoder-decoder frameworks often limit attention's effectiveness by applying it only in intermediate stages.

Purpose of the Study:

• To propose a novel deep learning model for 3D point cloud registration.
• To enhance the effectiveness of attention mechanisms in registration tasks.
• To address the computational cost and effectiveness trade-offs in attention-based registration.

Main Methods:

• Developed a novel model with attention layers integrated into both encoder and decoder stages.
• Employed self-attentional layers in the encoder to capture intra-point cloud relationships.
• Utilized cross-attentional layers in the decoder to enrich features with contextual information.

Main Results:

• The proposed model achieved high-quality results on a 3D point cloud registration task.
• Experiments on public datasets validated the model's effectiveness.
• Demonstrated superior performance compared to existing methods by leveraging integrated attention.

Conclusions:

• The novel model effectively improves 3D point cloud registration accuracy.
• Embedding attention in both encoder and decoder stages enhances feature representation and alignment.
• The approach offers a promising solution for computationally efficient and effective point cloud registration.