Learning Discriminative 3D Shape Representations by View Discerning Networks.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• View-based 3D shape recognition relies on extracting discriminative visual representations from projected images.
• Low-quality projections, especially with background clutter or object occlusion, significantly degrade 3D shape recognition performance.
• Existing methods struggle to adapt to varying projection quality in real-world scenarios.

Purpose of the Study:

• To propose a novel deep neural network, the View Discerning Network (VDN), for robust 3D shape recognition.
• To develop a mechanism that dynamically assesses and adjusts the contribution of different views to the final 3D shape representation.
• To enhance the discriminative power of visual features for 3D shape recognition under challenging conditions.

Main Methods:

• A novel deep neural network architecture, the View Discerning Network (VDN), is proposed.
• A Score Generation Unit (SGU) is devised to evaluate the quality of each projected image using score vectors.
• Two SGU structures, Channel-wise Score Unit and Part-wise Score Unit, are introduced to assess feature map quality from different perspectives.
• Features are weighted using the generated scores, and aggregated in an end-to-end framework to produce shape descriptors.

Main Results:

• The proposed View Discerning Network (VDN) significantly outperforms state-of-the-art methods in 3D shape retrieval tasks.
• Weighted features, informed by view quality scores, demonstrate superior performance compared to original features.
• The VDN exhibits excellent robustness against background clutter and object occlusion in real-world conditions.

Conclusions:

• The View Discerning Network (VDN) effectively addresses the challenge of low-quality projections in view-based 3D shape recognition.
• Dynamically adjusting feature contributions based on view quality enhances 3D shape representation and recognition accuracy.
• The VDN offers a robust and adaptable solution for 3D shape recognition in complex environments.