Multiple Discrimination and Pairwise CNN for view-based 3D object retrieval.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Analysis

Background:

• 3D object retrieval is a significant research area in computer vision.
• Deep learning features outperform traditional methods but existing networks neglect multi-view selection impacts.
• Contrastive loss alone inadequately addresses intra-class compactness and inter-class separability.

Purpose of the Study:

• To propose a novel Multi-view Discrimination and Pairwise CNN (MDPCNN) for improved 3D object retrieval.
• To address limitations in current deep learning approaches for multi-view 3D object retrieval.
• To enhance the discriminative power and retrieval performance of 3D object retrieval systems.

Main Methods:

• Developed MDPCNN incorporating Slice and Concat layers for simultaneous multi-batch and multi-view input.
• Employed clustering for training samples that are difficult to classify, enhancing network discriminability.
• Utilized a combination of contrastive-center loss and contrastive loss for optimized intra-class and inter-class feature representation.

Main Results:

• MDPCNN demonstrated superior performance in large-scale 3D object retrieval experiments.
• The proposed method achieved significant improvements over state-of-the-art algorithms.
• The integration of multi-view selection and advanced loss functions led to enhanced retrieval accuracy.

Conclusions:

• MDPCNN offers a significant advancement in 3D object retrieval technology.
• The approach effectively improves intra-class compactness and inter-class separability.
• This work provides a robust framework for future research in multi-view 3D object retrieval.