Event sensors are bio-inspired devices that mimic the brain’s efficient event-driven communication mechanisms. Unlike conventional sensors, which capture the scene synchronously at a fixed rate, event sensors report changes in the scene asynchronously.
For example, event sensors, such as DVS cameras, capture changes in brightness over time for each pixel independently rather than intensity images, as conventional sensors do. The benefits of higher temporal resolution, lower temporal latency, higher dynamic range, and higher power efficiency have sparked interest in machine learning for event data.
With the performance of deep learning (DL) algorithms improving on various tasks in recent years, he has focused on learning with event sensors. However, overfitting results in the degradation of the DL model which performs well on training data when validated against new and unknown data.
A simple solution to the overfitting problem is to significantly increase the amount of tagged data. This is, however, theoretically feasible but may be prohibitively expensive in practice. In addition, the severity of the overfitting problem increases in the event data due to its small size.
Data augmentation: a solution for generalization
Numerous studies suggest that increasing data improves the generalizability of DL models by increasing the amount and diversity of data from the available data. Translating, rotating, flipping, cropping, contrasting, sharpening, and other image enhancement techniques are quite common. However, event data is fundamentally different from frame-type data (like images), so we cannot directly apply the augmentation techniques developed for frame-type data to asynchronous event data.
Researchers from Chongqing University, National University of Singapore, German Aerospace Center and Tsinghua University present a new technique called EventDrop to increase event data by removing events. EventDrop is the first job to augment asynchronous event data by removing events in a way that is easy to implement, low in compute cost, and applicable to a variety of event-based tasks.
This study was inspired by observations that the number of events in a scene changes significantly over time. For example, the output of event cameras, for example, for the same scene under the same lighting conditions, can vary considerably over time. This may be due to noise from event cameras. It is possible to improve the diversity of event data and therefore the performance of downstream applications by randomly suppressing a proportion of events.
In addition, when performing certain tasks on real world data, the scenes of the images processed by the DL algorithms may be partially occluded. The ability of algorithms to generalize well to different datasets therefore strongly depends on the diversity of the training dataset in terms of occlusion. However, the majority of the available training data sets have low occlusion variance.
EventDrop addresses these issues by deleting selected events using various strategies to increase the diversity of training data. The researchers offer the following three methods for deciding which events should be suppressed:
- Random drop: it randomly removes a proportion of events in the sequence to overcome noise from event sensors.
- Drop by time: it removes events that occur at random time intervals, stimulating the case where objects are partially occluded for a specific time.
- Drop by area: It eliminates events that occur in a randomly chosen area of pixels while trying to improve data diversity by simulating various scenarios where some parts of objects are partially obscured.
These increase operations make it possible to increase the quantity of training data and the diversity of the data.
The researchers used the N-Caltech101 and N-Cars datasets to test EventDrop. They found that by removing events, their method could dramatically improve the accuracy of different deep neural networks on object classification tasks in the two datasets.
The team plans to apply the proposed method to other event-based learning tasks, including location reconnaissance, pose estimation, traffic flow estimation, simultaneous locating and mapping. , and many others.