Domain randomization for neural network classification | JoVE Visualize

Area of Science:

Computer Science
Artificial Intelligence
Machine Learning

Background:

Training neural networks, especially convolutional neural networks (CNNs), demands extensive labeled image datasets, often tens of thousands per category.
Acquiring and labeling these datasets is costly, time-consuming, and labor-intensive.
CNNs frequently struggle with generalization to out-of-domain test sets.

Purpose of the Study:

To investigate the efficacy of synthetic data generated via domain randomization (DR) for training neural network classifiers.
To determine the impact of various DR parameters on classifier accuracy and generalization.
To compare the performance of models trained on synthetic data against those trained on real-world data.

Main Methods:

Generated synthetic image datasets using domain randomization (DR) techniques.
Trained convolutional neural network (CNN) classifiers on the synthetic datasets.
Evaluated classifier performance on a baseline cats vs. dogs classification task and out-of-domain test sets.
Analyzed the significance of different DR parameters, including subject variety, lighting, and textures.

Main Results:

A well-generated synthetic dataset using DR achieved high accuracy (up to 88%) on a cats vs. dogs classification task, rivaling models trained on real datasets.
A wide variety of subjects was identified as the most crucial DR parameter for model accuracy.
Secondary parameters like lighting and textures had less impact on model performance.
Models trained on domain-randomized images demonstrated superior transfer learning capabilities to new domains compared to models trained on real photos.
Model performance remained stable with an increasing number of categories.

Conclusions:

Synthetic data generated through domain randomization offers a cost-effective and efficient alternative to large, manually labeled real-world datasets for training CNNs.
Domain randomization is a viable technique to improve the generalization ability of neural network classifiers.
Prioritizing subject variety in synthetic data generation is critical for maximizing classifier performance and out-of-domain transfer.