CSSE at WACV2027
Two papers from our chair have been accepted at the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV 2027). One grew out of Konstantin Kirchheim’s research stay at the University of Waterloo in Canada. The other introduces CarlAnomaly, a public dataset for testing how well automated vehicles can detect unexpected events
CarlAnomaly
A different challenge arises on the road. An unfamiliar obstacle may be visible in a single image, while a flickering streetlight only becomes apparent over time. CarlAnomaly: A Large-Scale Multimodal Anomaly Detection Benchmark for Autonomous Driving brings both kinds of events into a collection of 728 simulated driving scenarios with synchronized camera and LiDAR recordings. In the team’s experiments, different detectors often caught different events, even at the same false alarm rate. The dataset and evaluation tools are publicly available. CarlAnomaly project website
Scene Graph Reasoning
In Neuro-Symbolic Out-of-Distribution Detection via Scene Graph Reasoning, Konstantin Kirchheim worked with Prof. Czarnecki at Waterloo and Prof. Ortmeier in Magdeburg on a question many vision systems struggle to answer. What exactly makes an image look unusual? Their method uses a language model to help discover rules about the objects a camera sees and how they are arranged. When the system flags an image, researchers can inspect which expectations it violated. The approach worked best on images with recurring structures, such as faces and birds. Code on GitHub