🧠 Causal3D: A Benchmark for Visual Causal Reasoning

Causal3D is a dataset designed to evaluate models’ abilities to uncover latent causal relations from structured and visual data. This dataset integrates 3D-rendered scenes with tabular causal annotations, providing a unified testbed for advancing causal discovery, causal representation learning, and causal reasoning with vision-language models (VLMs) and large language models (LLMs).

🖼️ Visual Previews

Below are example images from different Causal3D scenes:

parabola	convex
magnetic	pendulum	reflection
seesaw	spring	water_flow

🗂️ Available Scenes

Below is the full list of builder configs you can load using load_dataset.

🔬 Hypothetical Scenes

Config Name	Description
`Hypothetical_V2_linear`	2 variables, linear causal relationship
`Hypothetical_V2_nonlinear`	2 variables, non-linear causal relationship
`Hypothetical_V3_fully_connected_linear`	3 variables, fully connected, linear
`Hypothetical_V3_v_structure_linear`	3 variables, V-structure, linear
`Hypothetical_V3_v_structure_nonlinear`	3 variables, V-structure, non-linear
`Hypothetical_V4_linear`	4 variables, linear causal relationship
`Hypothetical_V4_v_structure_nonlinear`	4 variables, V-structure, non-linear
`Hypothetical_V4_v_structure_linear`	4 variables, V-structure, linear
`Hypothetical_V5_linear`	5 variables, linear causal relationship
`Hypothetical_V5_v_structure_linear`	5 variables, V-structure, linear
`Hypothetical_V5_v_structure_nonlinear`	5 variables, V-structure, non-linear

🌍 Real-World Scenes

Config Name	Description
`Real_Parabola`	Real-world parabola trajectory
`Real_Magnet`	Real-world magnetic force
`Real_Spring`	Real-world spring oscillation
`Real_Water_flow`	Real-world water flow dynamics
`Real_Seesaw`	Real-world seesaw balance physics
`Real_Reflection`	Real-world light reflection
`Real_Pendulum`	Real-world pendulum motion
`Real_Convex_len`	Real-world convex lens refraction

🌐 Multi-View Real-World Scenes

Config Name	Description
`MV_Pendulum`	Multi-view real-world pendulum motion
`MV_H2_linear`	Multi-view H2 linear scene
`MV_H2_nonlinear`	Multi-view H2 nonlinear scene
`MV_H3_v_structure_linear`	Multi-view H3 V-structure linear scene
`MV_H4_fully_connected_linear`	Multi-view H4 fully connected linear scene
`MV_H4_v_structure_linear`	Multi-view H4 V-structure linear scene
`MV_H4_v_structure_nonlinear`	Multi-view H4 V-structure nonlinear scene
`MV_H5_fully_connected_linear`	Multi-view H5 fully connected linear scene
`MV_H5_v_structure_linear`	Multi-view H5 V-structure linear scene
`MV_H5_v_structure_nonlinear`	Multi-view H5 V-structure nonlinear scene

📚 Usage

🔹 Load from Hugging Face

You can easily load a specific scene using the Hugging Face datasets library:

from datasets import load_dataset

ds = load_dataset(
    "LLDDSS/Causal3D_Dataset",
    name="Real_Parabola",              # Replace with desired scene config name
    trust_remote_code=True             # Required for custom dataset loading
)

print(ds)

📌 Overview

While recent progress in AI and computer vision has been remarkable, there remains a major gap in evaluating causal reasoning over complex visual inputs. Causal3D bridges this gap by providing:

19 curated 3D-scene datasets simulating diverse real-world causal phenomena.
Paired tabular causal graphs and image observations across multiple views and backgrounds.
Benchmarks for evaluating models in both structured (tabular) and unstructured (image) modalities.

🎯 Evaluation Tasks

Causal3D supports a range of causal reasoning tasks, including:

Causal discovery from image sequences or tables
Intervention prediction under modified object states or backgrounds
Counterfactual reasoning across views
VLM-based causal inference given multimodal prompts

📊 Benchmark Results

We evaluate a diverse set of methods:

Classical causal discovery: PC, GES, NOTEARS
Causal representation learning: CausalVAE, ICM-based encoders
Vision-Language and Large Language Models: GPT-4V, Claude-3.5, Gemini-1.5

Key Findings:

As causal structures grow more complex, model performance drops significantly without strong prior assumptions.
A noticeable performance gap exists between models trained on structured data and those applied directly to visual inputs.