Reproduce “Hierarchical Question-Answering for Driving Scene Understanding Using Vision-Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Safaa Abdullahi Moallim Mohamud, Minjin Baek, Dong Seog Han
A hierarchical question-answering framework enhances compact vision-language models to provide detailed, real-time scene descriptions for autonomous driving, achieving a mean inference time of 423ms on a fine-tuned model with 94.81% accuracy.
Key insights
- A hierarchical question-answering framework enables systematic scene analysis for autonomous driving while significantly reducing computational overhead through dynamic question skipping.
- Fine-tuning compact Vision-Language Models (VLMs) on domain-specific, geographically-tailored datasets drastically improves scene understanding accuracy, reaching 94.81% compared to 72.10% for pre-trained models.
- Structured output generation, achieved by combining VLM capabilities with template-based post-processing, allows compact models to produce coherent, contextually accurate scene descriptions comparable to larger models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.