AnchorReasoning: A Visual Grounding and Causal Reasoning Dataset in Long-Tail Autonomous Driving Scenarios
AnchorReasoning provides visually grounded reasoning data for long-tail autonomous driving across 416,119 annotated frames.
AnchorReasoning is a visually grounded reasoning dataset for long-tail autonomous driving, built on WOD-E2E. It contains 416,119 annotated frames and 395,379 decision-critical elements across four major categories and 19 fine-grained types. Each frame is a visually grounded chain-of-thought covering element identification, attributes, action rationale, and trajectory planning. Curriculum fine-tuning across eight backbones cut 5-second ADE and FDE by 7.84 and 11.86 while using 18.5 fewer reasoning tokens and 0.32 seconds less inference per frame.
- Built on WOD-E2E with 416,119 frames and 395,379 decision-critical elements.
- Frames use visually grounded chain-of-thought linking evidence, rationale, and planning.
- Across eight backbones, 5-second ADE fell 7.84 and FDE fell 11.86.
- Models used 18.5 fewer reasoning tokens and 0.32 seconds less latency per frame.
From the original article
1 min read at arxiv.orgVision-language models (VLMs) offer a promising approach to long-tail autonomous driving, but existing driving datasets provide limited supervision for connecting decision-critical visual evidence with reasoning and planning.
By Zhipeng Bao · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.