3D2SeqViews: Aggregating Sequential Views for 3D Global Feature Learning by CNN With Hierarchical Attention...

Area of Science:

• Computer Vision
• Machine Learning
• 3D Shape Analysis

Background:

• Aggregating multiple views is crucial for learning 3D global features.
• Traditional pooling methods in deep learning overlook intra-view content and inter-view spatial relationships, limiting feature discriminability.
• Existing methods struggle to effectively capture the sequential nature of multi-view data.

Purpose of the Study:

• To propose a novel method, 3D to Sequential Views (3D2SeqViews), for more effective aggregation of sequential views in 3D shape analysis.
• To enhance the discriminability of learned 3D features by addressing limitations of conventional pooling techniques.
• To improve performance in 3D shape classification and retrieval tasks.

Main Methods:

• Encoding content information within each 3D view.
• Developing a hierarchical attention aggregation mechanism that simultaneously considers view content and sequential spatiality.
• Introducing view-level attention for weighting sequential views via recursive view integration, robust to initial view position.
• Incorporating class-level attention to leverage fine-tuned network discriminative abilities.

Main Results:

• 3D2SeqViews learns more discriminative 3D global features compared to state-of-the-art methods.
• The proposed method achieves superior performance in 3D shape classification and retrieval tasks.
• Outperforming results were demonstrated across three large-scale benchmarks.

Conclusions:

• The 3D2SeqViews method effectively aggregates sequential views using hierarchical attention, overcoming pooling limitations.
• This approach significantly enhances feature learning for 3D shape analysis.
• The method offers a promising advancement for 3D shape classification and retrieval applications.