[ICCV 2025] Spatial-Temporal Intelligence Benchmark (STI-Bench)

This repository contains the Spatial-Temporal Intelligence Benchmark (STI-Bench), introduced in the paper “STI-Bench: Are MLLMs Ready for Precise Spatial-Temporal World Understanding?”, which evaluates the ability of Multimodal Large Language Models (MLLMs) to understand spatial-temporal concepts through real-world video data.

Files

# Make sure git-lfs is installed (https://git-lfs.com)
git lfs install
git clone https://huggingface.co/datasets/MIRA-SJTU/STI-Bench

Dataset Description

STI-Bench evaluates MLLMs’ spatial-temporal understanding by testing their ability to estimate, predict, and understand object appearance, pose, displacement, and motion from video data. The benchmark contains more than 2,000 question-answer pairs across 300 videos, sourced from real-world environments such as desktop settings, indoor scenes, and outdoor scenarios. These videos are taken from datasets like Omni6DPose, ScanNet, and Waymo.

STI-Bench is designed to challenge models on both static and dynamic spatial-temporal tasks, including:

Task Name	Description
3D Video Grounding	Locate the 3D bounding box of objects in the video
Ego-Centric Orientation	Estimate the camera's rotation angle
Pose Estimation	Determine the camera pose
Dimensional Measurement	Measure the length of objects
Displacement & Path Length	Estimate the distance traveled by objects or camera
Speed & Acceleration	Predict the speed and acceleration of moving objects or camera
Spatial Relation	Identify the relative positions of objects
Trajectory Description	Summarize the trajectory of moving objects or camera

Dataset Fields Explanation

The dataset contains the following fields, each with its respective description:

Field Name	Description
Video	The string corresponding to the video file.
Source	The string corresponding to the video source, which can be "ScanNet," "Waymo," or "Omni6DPose."
Task	The string representing the task type, e.g., "3D Video Grounding," "Ego-Centric Orientation," etc.
QType	The string specifying the question type, typically a multiple-choice question.
Question	The string containing the question presented to the model.
Prompt	Additional information that might be helpful for answering the question, such as object descriptions.
time_start	A float64 value indicating the start time of the question in the video (in seconds).
time_end	A float64 value indicating the end time of the question in the video (in seconds).
Candidates	A dictionary containing answer choices in the format {"A": "value", "B": "value", ...}.
Answer	The string corresponding to the correct answer, represented by the choice label (e.g., "A", "B", etc.).
Answer Detail	A string representing the precise value or description of the correct answer.
ID	A sequential ID for each question, unique within that video.
Scene	The string describing the scene type of the video, such as "indoor," "outdoor," or "desktop."

Evaluation

STI-Bench evaluates performance using accuracy, calculated based on exact matches for multiple-choice questions.

We provide an out-of-the-box evaluation of STI-Bench in our GitHub repository

Note: STI-Bench is now supported by EASI, an open-source evaluation tool for spatial intelligence.

Citation

@article{li2025sti,
    title={STI-Bench: Are MLLMs Ready for Precise Spatial-Temporal World Understanding?}, 
    author={Yun Li and Yiming Zhang and Tao Lin and XiangRui Liu and Wenxiao Cai and Zheng Liu and Bo Zhao},
    year={2025},
    journal={arXiv preprint arXiv:2503.23765},
}