EventDrop: data augmentation for event-based learning

The advantages of event-sensing over conventional sensors (e.g., higher dynamic range, lower time latency, and lower power consumption) have spurred research into machine learning for event data. Unsurprisingly, deep learning has emerged as a competitive methodology for learning with event sensors; in typical setups, discrete and asynchronous events are first converted into frame-like tensors on which standard deep networks can be applied. However, over-fitting remains a challenge, particularly since event datasets remain small relative to conventional datasets (e.g., ImageNet). In this paper, we introduce EventDrop, a new method for augmenting asynchronous event data to improve the generalization of deep models. By dropping events selected with various strategies, we are able to increase the diversity of training data (e.g., to simulate various levels of occlusion). From a practical perspective, EventDrop is simple to implement and computationally low-cost. Experiments on two event datasets (N-Caltech101 and N-Cars) demonstrate that EventDrop can significantly improve the generalization performance across a variety of deep networks.

Camera Frame Misalignment in a Teleoperated Eye-in-Hand Robot: Effects and a Simple Correction Method

Misalignment between the camera frame and the operator frame is commonly seen in a teleoperated system and usually degrades the operation performance. The effects of such misalignment have not been fully investigated for eye-in-hand systems - systems that have the camera (eye) mounted to the end-effector (hand) to gain compactness in confined spaces such as in endoscopic surgery. This paper provides a systematic study on the effects of the camera frame misalignment in a teleoperated eye-in-hand robot and proposes a simple correction method in the view display. A simulation is designed to compare the effects of the misalignment under different conditions. Users are asked to move a rigid body from its initial position to the specified target position via teleoperation, with different levels of misalignment simulated. It is found that misalignment between the input motion and the output view is much more difficult to compensate by the operators when it is in the orthogonal direction (~40s) compared with the opposite direction (~20s). An experiment on a real concentric tube robot with an eye-in-hand configuration is also conducted. Users are asked to telemanipulate the robot to complete a pick-and-place task. Results show that with the correction enabled, there is a significant improvement in the operation performance in terms of completion time (mean 40.6%, median 38.6%), trajectory length (mean 34.3%, median 28.1%), difficulty (50.5%), unsteadiness (49.4%), and mental stress (60.9%).