SeqViews2SeqLabels: Learning 3D Global Features via Aggregating Sequential Views by RNN With Attention.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Shape Analysis

Background:

• Current deep learning models for 3D shape analysis aggregate multiple views using pooling methods.
• Pooling disregards content and spatial information among views, limiting feature learning.
• Existing methods struggle with limited 3D shape datasets, leading to overfitting.

Purpose of the Study:

• To propose a novel deep learning model, SeqViews2SeqLabels, for enhanced 3D shape analysis.
• To overcome the limitations of pooling in view aggregation for 3D shape classification.
• To improve the discriminative ability of learned global features for 3D shapes.

Main Methods:

• Developed SeqViews2SeqLabels, an encoder-decoder model utilizing recurrent neural networks (RNNs) with attention.
• The encoder-RNN learns global features by encoding spatial and content information from sequential views.
• The decoder-RNN performs classification by predicting sequential labels, enhanced by an attention mechanism.

Main Results:

• SeqViews2SeqLabels effectively aggregates sequential views, capturing semantics and improving global feature learning.
• The model demonstrates superior performance in 3D shape classification and retrieval across large-scale benchmarks.
• The attention mechanism enhances feature distinctiveness and mitigates the impact of initial view selection.

Conclusions:

• SeqViews2SeqLabels offers a more effective approach to view aggregation for 3D shape analysis compared to pooling.
• The proposed method learns more discriminative global features, leading to state-of-the-art results.
• Predicting sequential labels and using attention alleviates overfitting and improves classification accuracy.