Reproduce “How Real is CARLAs Dynamic Vision Sensor? A Study on the Sim-to-Real Gap in Traffic Object Detection” 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
Kaiyuan Tan, Pavan Kumar B N, Bharatesh Chakravarthi
This study quantifies the sim-to-real gap for CARLA's Dynamic Vision Sensor (DVS) module in traffic object detection, demonstrating that models trained purely on synthetic data perform poorly on real-world test sets. Performance on real-world data improves almost linearly as the proportion of real-world data in the training set is increased.
Key insights
- A substantial sim-to-real gap exists between CARLA's DVS-generated events and real event camera data for traffic object detection.
- Integrating real-world data into the training process is crucial, as model performance on real-world test sets improves nearly linearly with an increasing proportion of real data.
- The current fidelity of CARLA’s DVS simulation has limitations, particularly under specific conditions like nighttime, which poses challenges for generalization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.