Generative Model With Coordinate Metric Learning for Object Recognition Based on 3D Models.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Data acquisition and labeling are significant challenges in deep learning.
• Generating realistic synthetic data can mitigate the need for extensive real-world data collection.

Purpose of the Study:

• To propose a generative model that utilizes synthetic images from 3D models to improve deep learning database creation.
• To reduce the burden of collecting and labeling real-world training data.
• To enhance the realism of background conditions in training datasets.

Main Methods:

• A two-sub-network architecture: a semantic foreground object reconstruction network (Bayesian inference) and a classification network (multi-triplet cost training).
• Metric learning with additional foreground object channels for image recognition.
• Pose-based multi-triplet cost function for training classifiers on synthetic data.
• A coordinate training strategy with adaptive noise to synchronize sub-network convergence.

Main Results:

• Achieved state-of-the-art accuracy of 50.5% on the ShapeNet database.
• Demonstrated successful data migration from synthetic to real images.
• The pipeline enables real-image recognition solely based on 3D models.

Conclusions:

• The proposed generative model effectively reduces the challenges of data collection for deep learning.
• The method allows for training robust classifiers using only synthetic data.
• This approach offers a viable pathway for object recognition in real-world scenarios using 3D model data.