Domain-Adversarial-Guided Siamese Network for Unsupervised Cross-Domain 3-D Object Retrieval

Area of Science:

• Computer Vision
• Machine Learning
• 3-D Data Analysis

Background:

• Massive 3-D data availability from advanced sensors necessitates efficient processing.
• Manual labeling of 3-D objects is time-consuming and impractical for large datasets.
• Existing cross-domain 3-D object retrieval methods often struggle with feature abstraction and domain shift.

Purpose of the Study:

• To develop an unsupervised method for cross-domain 3-D object retrieval (CD3DOR).
• To enable knowledge transfer from labeled 2-D images or 3-D objects to unlabeled 3-D objects.
• To enhance feature abstraction capabilities beyond simple domain shift elimination.

Main Methods:

• A domain-adversarial guided siamese network (DAGSN) is proposed.
• Siamese networks are used for balanced accuracy and efficiency in encoding 3-D objects and 2-D images.
• Mutual information (MI) maximization enhances feature abstraction, while a conditional domain classifier aligns features across domains.

Main Results:

• The DAGSN generates domain-invariant and discriminative features for effective CD3DOR.
• Experiments on cross-dataset (3-D to 3-D) and cross-modal (2-D to 3-D) retrieval show significant performance improvements.
• The proposed method outperforms existing state-of-the-art CD3DOR techniques.

Conclusions:

• DAGSN effectively addresses the challenge of unsupervised cross-domain 3-D object retrieval.
• The integration of MI-based enhancement and conditional domain classification is key to the method's success.
• This approach offers a robust solution for leveraging labeled data to improve retrieval from large, unlabeled 3-D datasets.