Distance-Annotated Traffic Perception Question Answering (DTPQA)
Description
Distance-Annotated Traffic Perception Question Answering (DTPQA) is a Visual Question Answering (VQA) benchmark designed specifically for this purpose: it can be used to evaluate the perception systems of VLMs in traffic scenarios using trivial yet crucial questions relevant to driving decisions. It consists of two parts: a synthetic benchmark (DTP-Synthetic) created using a simulator, and a real-world benchmark (DTP-Real) built on top of existing images of real traffic scenes. Additionally, DTPQA includes distance annotations, i.e., how far the object in question is from the camera. More specifically, each DTPQA sample consists of (at least): (a) an image, (b) a question, (c) the ground truth answer, and (d) the distance of the object in question, enabling analysis of how VLM performance degrades with increasing object distance
Files
Institutions
- University of Limerick