In recent years, dynamic vision sensors (DVS), also known as event-based cameras or neuromorphic sensors, have seen increased use due to various advantages over conventional frame-based cameras. Using principles inspired by the retina, its high temporal resolution overcomes motion blurring, its high dynamic range overcomes extreme illumination conditions and its low power consumption makes it ideal for embedded systems on platforms such as drones and self-driving cars. However, event-based data sets are scarce and labels are even rarer for tasks such as object detection. We transferred discriminative knowledge from a state-of-the-art frame-based convolutional neural network (CNN) to the event-based modality via intermediate pseudo-labels, which are used as targets for supervised learning. We show, for the first time, event-based car detection under ego-motion in a real environment at 100 frames per second with a test average precision of 40.3% relative to our annotated ground truth. The event-based car detector handles motion blur and poor illumination conditions despite not explicitly trained to do so, and even complements frame-based CNN detectors, suggesting that it has learnt generalized visual representations.

Proton radiography is a technique extensively used to resolve magnetic field structures in high-energy-density plasmas, revealing a whole variety of interesting phenomena such as magnetic reconnection and collisionless shocks found in astrophysical systems. Existing methods of analyzing proton radiographs give mostly qualitative results or specific quantitative parameters, such as magnetic field strength, and recent work showed that the line-integrated transverse magnetic field can be reconstructed in specific regimes where many simplifying assumptions were needed. Using artificial neural networks, we demonstrate for the first time 3D reconstruction of magnetic fields in the nonlinear regime, an improvement over existing methods, which reconstruct only in 2D and in the linear regime. A proof of concept is presented here, with mean reconstruction errors of less than 5% even after introducing noise. We demonstrate that over the long term, this approach is more computationally efficient compared to other techniques. We also highlight the need for proton tomography because (i) certain field structures cannot be reconstructed from a single radiograph and (ii) errors can be further reduced when reconstruction is performed on radiographs generated by proton beams fired in different directions.

No abstract