Syntethic Data in Computer Vision
By Pierluigi Salvo Rossi
Convolutional neural networks (CNNs) and related variations have significantly impacted the areas of image processing and computer vision. Amazing performances, superior to human capability, have been achieved in several tasks (e.g. object detection and classification) with affordable computational complexity.
The main price is paid in terms of the huge amount of labeled data needed for training. While for some specific scenario satisfactory datasets may be easily available, in general this is not always the case. Building datasets for some scenarios of interest may be severely challenging due to the intrinsic difficulty for image acquisition (e.g. safety issues, security issues, economical issues) or time-consuming labeling effort.
Synthetic data generated from computer graphics may provide easy and fast acquisition and labeling solutions, thus representing an alternative and effective strategy to fuel computer-vision approaches based on big-data technology with relevant training datasets. However, synthetic data suffers from dataset bias, making models trained on synthetic data typically underperform in real-world scenarios.
State-of-the-art domain-adaptation techniques (e.g. fine tuning, layer freezing, data augmentation) enable the use of synthetic data in computer-vision applications when real-world data is scarce. Kongsberg Digital is currently exploring these techniques in order to exploit the various simulators developed in different domains. More specifically, as for collaboration with NTNU, computer-vision applications in the aquaculture and maritime domains have been prioritized. Kongsberg Digital’s simulators will be briefly introduced and ongoing activities within domain adaptation using synthetic data will be illustrated.
Pierluigi Salvo Rossi
Advanced Analytics & Machine Learning Group Kongsberg Digital AS, Norway