A successfull deep learning based system relies on being trained on a large and representative dataset which can be costly to acquire. We are also working on methods to create realistic looking labeled datasets based on simulated data, which can be created more cost-efficiently.
Our expertise
To further develop and link novel technologies and methodologies within automation to support innovation processes and advanced work systems in the manufacturing industries. First case is developing a flexible and robust bin-picking system based on deep learning. The convolutional neural network is trained in a simulation environment and deployed on real 3D data.
This work is done in the project:
Deep learning requires large annotated datasets which can be difficult and costly to acquire. We are developing a framework that exploits generative adversarial networks (GAN) for creating realistic training datasets based on simulated and unsupervised images. Images (RGB-D) are simulated by a rendering process together with relevant annotation information. A model is learnt to improve the realism of the simulated images using realistic unannotated images while preserving the simulated annotation information. This model allow us to very cost efficient generate large realistic annotated datasets suitable for training deep-learning networks.
Normality" focused pipeline for understanding ROV and diver video from fish cages. Data driven detection of regions containing fish-cage structures (moorings, ropes etc) and defects (holes, fouling). Semantic classification of extracted regions.
This work is done in the project:
