Abstract

Natural Language (NL) descriptions can be the most convenient or the only wayto interact with systems built to understand and detect city scale trafficpatterns and vehicle-related events. In this paper, we extend the widelyadopted CityFlow Benchmark with natural language descriptions for vehicletargets and introduce the CityFlow-NL Benchmark. The CityFlow-NL contains morethan 5,000 unique and precise NL descriptions of vehicle targets, making it thelargest-scale tracking with NL descriptions dataset to our knowledge. Moreover,the dataset facilitates research at the intersection of multi-object tracking,retrieval by NL descriptions, and temporal localization of events.

 

Quick Read (beta)