OmniTraffic: A Controllable Generation Pipeline and Benchmark for Spatio-Temporal Traffic Reasoning
Researchers have introduced OmniTraffic, a new pipeline and benchmark designed to improve spatio-temporal reasoning in traffic scenarios. This system utilizes 3D reconstructed environments and real-world surveillance footage to generate a large dataset of question-answering samples focused on traffic perception, multi-view, and temporal reasoning. Evaluations of current large multimodal models (LMMs) on OmniTraffic revealed a significant gap between human and model performance, particularly in topology-grounded and spatio-temporal tasks. The study also demonstrated that fine-tuning LMMs on simulated OmniTraffic data can enhance their performance on real-world traffic scenes. AI