Reproduce “Predicting Driver Attention in Critical Situations” 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
Ye Xia, Danqing Zhang, Jinkyu Kim, Ken Nakayama, Karl Zipser, David Whitney
Researchers at the University of California, Berkeley, developed a deep neural network model for predicting driver attention in critical situations, utilizing a novel in-lab data collection protocol to create the BDD-A dataset and a Human Weighted Sampling method. The model demonstrates superior accuracy in identifying human-like attention, especially in hazardous scenarios, and generalizes effectively from lab-collected data to real-world driving conditions.
Key insights
- Aggregating independent human gazes in a lab environment effectively filters out noise and captures driving-relevant multi-focus attention, offering higher quality data than single-driver in-car recordings.
- A training strategy that weights samples based on the 'uncommonness' of human attention patterns (Human Weighted Sampling) can successfully guide models to learn from critical, rare events.
- Data collected efficiently in a lab setting can be robustly generalized to predict real-world in-car driver attention, providing a scalable solution for hard-to-collect critical scenario data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.