QuantiPhy

Dataset Summary

QuantiPhy is a benchmark for evaluating whether vision–language models (VLMs) can perform quantitative physical inference from visual evidence, rather than producing plausible but ungrounded numerical guesses.

This repository contains the official test set of the QuantiPhy benchmark, consisting of 3,373 video–question (QA) pairs across 556 videos. Ground-truth answers are withheld to ensure fair evaluation.

Each instance requires a model to output a single continuous numerical value (e.g., object size, velocity, or acceleration) in real-world units, given a short video and a natural-language question.

Looking for the validation set?
A 159-sample validation split with ground-truth answers is available at PaulineLi/QuantiPhy-validation for model development, prompt tuning, and ablation studies.

---

Supported Tasks

• Video-based numerical regression
• Quantitative visual reasoning
• Vision–language model evaluation

Tasks cover three core kinematic properties:

• Size
• Velocity
• Acceleration

All questions are open-ended and require predicting a real-valued scalar.

---

Dataset Structure

Each instance is represented as a structured video–text record with the following fields:

Field	Description
video_id	Unique identifier for the video (maps to <video_id>.mp4 in the video folders)
video_source	Data source (simulation, lab, internet, or segmentation)
video_type	Four-character code encoding task configuration (see below)
fps	Frame rate of the video
inference_type	Prior/target configuration: SS, SD, DS, or DD
question	Natural-language question with explicit physical units
prior	Physical prior provided in world units (e.g., object size, velocity, or acceleration)
depth_info	Depth/distance information for 3D configurations (null for 2D tasks)

Videos are short (typically 2–3 seconds) and recorded with a static camera to ensure well-defined kinematic inference.

Video Type Code

Each video_type is a 4-character code [P][D][O][B]:

Position	Meaning	Values
P — Physical prior	S = Size, V = Velocity, A = Acceleration
D — Dimensionality	2 = 2D (planar), 3 = 3D (with depth)
O — Object setting	S = Single-object, M = Multi-object
B — Background	X = Plain, S = Simple, C = Complex

Inference Type

Code	Prior	Target	Description
SS	Static	Static	Infer a static quantity from a static prior
SD	Static	Dynamic	Infer a dynamic quantity from a static prior
DS	Dynamic	Static	Infer a static quantity from a dynamic prior
DD	Dynamic	Dynamic	Infer a dynamic quantity from a dynamic prior

---

Task Design Overview

Each instance provides the model with:

• a short video depicting object motion, and
• one physical prior in world units (object size, velocity at a given timestamp, or acceleration at a given timestamp).

The model is then asked to infer a target kinematic quantity—possibly for a different object—expressed in real-world units.

Tasks vary along four axes:

1. Physical prior: Size (S), Velocity (V), Acceleration (A)
2. Dimensionality: 2D (planar motion) or 3D (with depth variation)
3. Object setting: Single-object (S) or multi-object (M)
4. Background complexity: Plain (X), Simple (S), Complex (C)

---

Dataset Statistics

	Count
QA pairs	3,373
Unique videos	556
Video types	37 (18 × 2D + 19 × 3D)

By source:

Source	QA pairs
Simulation	1,633
Lab	811
Internet	547
Segmentation	382

By inference type:

Type	QA pairs
DS (Dynamic → Static)	1,689
SS (Static → Static)	649
SD (Static → Dynamic)	585
DD (Dynamic → Dynamic)	450

---

Videos

Videos are provided in two resolutions:

• quantiphy_fullset_videos/ — original resolution
• quantiphy_fullset_videos_480p/ — 480p (for faster download and lower-resolution evaluation)

Each video filename corresponds to the video_id field in the dataset (e.g., simulation_0007.mp4).

---

Usage

from datasets import load_dataset

ds = load_dataset("PaulineLi/QuantiPhy", split="test")
print(ds[0])
# {'video_id': 'simulation_0007', 'video_source': 'simulation', ...}

For the validation set with ground-truth answers:

ds_val = load_dataset("PaulineLi/QuantiPhy-validation", split="validation")

---

Data Sources and Quality Control

• Simulation: Blender-rendered scenes with precise physical ground truth.
• Laboratory capture: Real-world recordings using calibrated depth and multi-view setups.
• Internet / author-recorded videos: Carefully curated monocular videos meeting strict physical constraints.
• Segmentation: Videos with segmented objects for controlled evaluation.

All videos undergo manual review to remove:

• excessive motion blur,
• severe occlusion,
• untrackable motion,
• personally identifiable information (PII).

---

License

The annotations and metadata in this repository are released under the
Creative Commons Attribution 4.0 (CC BY 4.0) license.

Videos originate from simulated environments, laboratory recordings, and publicly available sources.
Each video remains subject to its original license and terms of use.

This release is intended for research and evaluation purposes.

---

Authors

Puyin Li*, Tiange Xiang*, Ella Mao*,
Shirley Wei, Xinye Chen, Adnan Masood,
Li Fei-Fei†, Ehsan Adeli†

* Equal contribution.

---

Citation

If you use QuantiPhy in your work, please cite:

@article{li2025quantiphy,
  title   = {QuantiPhy: A Quantitative Benchmark Evaluating Physical Reasoning Abilities of Vision-Language Models},
  author  = {Li, Puyin and Xiang, Tiange and Mao, Ella and Wei, Shirley and Chen, Xinye and Masood, Adnan and Li, Fei-Fei and Adeli, Ehsan},
  journal = {arXiv preprint arXiv:2512.19526},
  year    = {2025}
}