Reproduce “DSEC: A Stereo Event Camera Dataset for Driving Scenarios” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Mathias Gehrig, Willem Aarents, Daniel Gehrig, Davide Scaramuzza
DSEC is a multimodal driving dataset designed to support stereo depth research with event cameras. It contains 53 synchronized sequences spanning day and night, pairing high-resolution stereo event streams with RGB images, lidar, RTK-GPS, and disparity labels; its baseline results also show that a stereo model trained on MVSEC transfers poorly to DSEC without retraining.
Key insights
- The event stereo pair has a 60 cm baseline, and the dataset includes both daylight and night driving, enabling evaluation across varied illumination and sensor conditions.
- Cross-dataset transfer is difficult: the MVSEC-trained baseline has an 85.9% D1 error on DSEC, while DSEC retraining reduces the reported overall D1 error to about 11%.
- Disparity labels are useful but incomplete: filtering and lidar coverage make them sparser on moving objects and in challenging conditions, and night performance is affected by lower signal-to-noise ratio.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.