Contrast Thresholding and Subsampling: Evaluating Event Reduction Strategies for Automotive Applications

Event Bias vs. Subsampling:

This deposit contains the code used to study whether reducing the event rate of an event camera at the sensor (by raising the contrast threshold, or "bias") is equivalent to reducing it afterwards by subsampling a dense event stream. The two are compared on a downstream task: object detection with a Recurrent Vision Transformer (RVT).

It includes two components: a library of event subsampling methods with tooling to apply them across a dataset and to match event rates between variants, and an adapted version of the RVT object detector used to train a model on each event stream variant and cross-evaluate every model on every variant.