RotationNet for Joint Object Categorization and Unsupervised Pose Estimation from Multi-View Images.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Object pose and category estimation are crucial in computer vision.
• Existing methods often require labeled viewpoint data for training.
• Handling datasets with unaligned object views presents a significant challenge.

Purpose of the Study:

• To develop an unsupervised method for joint object category and pose estimation using multi-view images.
• To enable accurate inference even with partial object views.
• To create shared, view-specific feature representations across object classes.

Main Methods:

• A Convolutional Neural Network (CNN)-based model named RotationNet was proposed.
• Viewpoint labels were treated as latent variables, learned in an unsupervised manner.
• A pose alignment strategy was employed to generate shared feature representations.

Main Results:

• RotationNet demonstrated superior performance in 3D object classification on ModelNet datasets compared to state-of-the-art methods.
• The model achieved comparable performance in object pose estimation without prior pose knowledge.
• RotationNet-based object ranking won first prize in two tracks of the 3D Shape Retrieval Contest (SHREC) 2017.

Conclusions:

• RotationNet effectively performs unsupervised joint object category and pose estimation from multi-view images.
• The model's ability to use partial views and shared features enhances practical applicability and accuracy.
• The method shows strong potential for real-world applications, including 3D shape retrieval and analysis.