Trace4D: Grounded Spatiotemporal Reasoning via 4D State Sequences

概要

Understanding how objects and their spatial relationships evolve in 3D space over time remains challenging for vision-language models (VLMs), particularly under changing viewpoints. Recent approaches incorporate geometric information into reasoning, but explicit supervision for temporally ordered 4D state evolution and its use as intermediate evidence remains limited. We introduce \textbf{Trace4D}, which trains VLMs to predict 4D state sequences and reason over their inter-timestep changes alongside temporal visual observations. Its 4D-grounded chain-of-thought (CoT) explicitly represents 4D states, analyzes temporal visual context, and connects both as evidence for the answer. We construct Trace4D-Train from in-the-wild videos with 600K 4D perception examples and approximately 70K grounded QA examples, and train with joint supervised fine-tuning followed by geometry-aware GRPO. At inference, Trace4D predicts 4D states and reasoning directly from video without reference annotations or an external geometric model. Experiments on Trace4D-Bench and VLM4D show improved QA accuracy over the base VLMs and benefits from reasoning over 4D state changes and temporal visual observations.

参考リンク

• Paper ↗

リンク先のライセンス・再現性・データ条件をご確認ください。自動分類や実装の目安は保証ではありません。

実装を始めるプロンプト

参考リンクを確認したうえで、次の内容をコピーして利用できます。

# 論文・記事
Trace4D: Grounded Spatiotemporal Reasoning via 4D State Sequences

# 概要
Understanding how objects and their spatial relationships evolve in 3D space over time remains challenging for vision-language models (VLMs), particularly under changing viewpoints. Recent approaches…

# 原典
https://arxiv.org/abs/2611.02759

上記の手法を実装してください。実行前に依存関係とライセンスを確認してください。