This paper compares four vision-language models—Phi-3, Qwen2-VL, LLaVA, and PaliGemma—on the RoadscapesQA dataset for road scene question answering. Three task types were evaluated: object counting, object description, and surrounding description, each tested under daytime and nighttime conditions separately. A lightweight inference pipeline was built using prompt-based evaluation, answer normalization, and per-category accuracy reporting. Qwen2-VL achieved the highest overall accuracy, followed by Phi-3, LLaVA, and PaliGemma. The results show that daytime scenes are generally easier than nighttime scenes, while object description remains the most difficult task across all models. This work provides a small, reproducible benchmark-style comparison for evaluating vision-language model behavior on Indian road-scene data.
Vishal (Sat,) studied this question.