Abstract
Successful deployment of machine learning models for computer vision on consumer devices requires generalization across multiple devices with varying camera sensor characteristics and image signal processing (ISP) tuning profiles. To ensure model performance, expensive data collection campaigns need to be executed to capture device-specific data for various devices to obtain training datasets to train such machine learning models. This disclosure describes techniques to generate training data to train a machine learning model for deployment across multiple target devices. Variations in camera sensor and ISP tuning profiles are simulated using randomized RAW transformations to produce different color sensitivities and by randomizing ISP tuning profiles to generate different rendering styles. Synthetic training data thus obtained covers a wide range of hypothetical target devices with varying camera sensor characteristics and tuning profiles, and can be used to train a machine learning model to generate robust feature representations that generalize to a wide range of devices, enabling zero-shot generalization to unknown future devices without the need for additional device-specific data collection.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Abdelkader, Ahmed and Chu, Wen-Sheng, "Generation of Synthetic Training Data Using Randomization in Camera Sensor Response and Image Signal Processing Pipeline", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/11390