Reproduce “Real-World On-Vehicle Evaluation of Embedding-Based Anomaly 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
Albert Schotschneider, Daniel Bogdoll, Svetlana Pavlitska, Ahmed Abouelazm, Johann Marius Zoellner
Researchers from FZI Research Center for Information Technology and KIT Karlsruhe Institute of Technology developed and evaluated a minimalist, training-free anomaly detection method using pretrained DINOv3 embeddings for autonomous vehicles. The system, deployed on a real autonomous vehicle, achieved real-time inference rates of 12.5 Hz and consistently localized various anomalies in dynamic urban and rural driving environments.
Key insights
- Pretrained vision transformer embeddings (DINOv3) provide sufficiently rich and generalizable semantic features for robustly detecting anomalies in complex, real-world autonomous driving scenes.
- A remarkably simple setup, using only a single reference image, can effectively define normal scene conditions and enable practical, real-time anomaly detection without explicit training on anomalous data.
- Embedding-based anomaly detection methods are feasible for real-time deployment on physical autonomous vehicles, offering localized anomaly information crucial for downstream perception and planning modules.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.