DeLux Training Dataset

This dataset supports research on detecting and removing lighting artifacts (flare, glare, overexposure, and flicker) from RGB video by leveraging neuromorphic event data. 

The complete dataset is organized into two top-level directories: removal and detection. The removal directory is divided into the subdirectories train, val, and test, each containing preprocessed samples in the .npy format. In the train subset, each sample is represented as a 5-channel array: the first three channels correspond to the RGB frame, the fourth contains the grayscale event reconstruction, and the fifth is a binary mask that indicates where the event data were available (with value 255). The val and test subsets extend this to 8-channel samples, where the final three channels represent the RGB values of the synthetic lighting artifact overlaid on the input. In addition, the removal directory contains two files: test artifact source.json and val artifact source.json, which map the samples from test and validation subsets to the artifact generators.

The detection directory contains two subdirectories: flare7k and event. Each sample in these sets is a 4-channel array, with the first three channels encoding the RGB frame and the fourth channel containing the binary segmentation mask of artifact regions. The flare7k directory includes samples derived from the Flare7k++ dataset, while the event directory includes our manually annotated real-world samples from the DSEC and E2VID recordings.