SINTEF has a long track record of developing state-of-the-art deep learning based methods for industrial partners to increase the value of their data for e.g., prediction, detection or classification. Typically, the task is to extract useful information out of signals buried within vast amounts of possibly noisy and otherwise imperfect data coming from heterogenous data-streams.
We work primarily on developing deep learning based systems within these areas:
• Computer Vision (2D/3D images and video)
• Acoustics
• Healthcare
• Industrial sensor data
• General prediction and detection systems
For deep learning-based systems to be applicable for industrial applications, the inference needs to be robust, accurate and precise. This typically involves supervised training on large amounts of structured and hand-labelled data is generally very costly to generate for many of our customers. A primary focus of our research is to help our customers develop deep learning-based systems which can be efficiently trained even without costly hand-labelled datasets. Examples of techniques that we develop are:
• Self-supervised training
• Unsupervised training
• Training on simulated data and closing the reality gap
• Domain informed or guided deep learning
Here is an assorted selection of problems we have successfully tackled with deep learning:
• Detect fish species and estimate fish size
• Semantic segmentation and labeling of large-scale 3D point clouds
• Robotic bin-picking
• Automatic 6-DOF localization of autonomous underwater vehicles
• Electricity market bidding strategy selection