Spatial457: A Diagnostic Benchmark for 6D Spatial Reasoning of Large Multimodal Models

Xingrui Wang¹, Wufei Ma¹, Tiezheng Zhang¹, Celso M. de Melo², Jieneng Chen¹, Alan Yuille¹

¹ Johns Hopkins University      ² DEVCOM Army Research Laboratory

🌐 Project Page • 📄 Paper • 🤗 Dataset • 💻 Code

---

🧠 Introduction

Spatial457 is a diagnostic benchmark designed to evaluate 6D spatial reasoning in large multimodal models (LMMs). It systematically introduces four core spatial capabilities:

• 🧱 Multi-object understanding
• 🧭 2D spatial localization
• 📦 3D spatial localization
• 🔄 3D orientation estimation

These are assessed across five difficulty levels and seven diverse question types, ranging from simple object queries to complex reasoning about physical interactions.

---

📂 Dataset Structure

The dataset is organized as follows:

Spatial457/
├── images/                     # RGB images used in VQA tasks
├── questions/                  # JSONs for each subtask
│   ├── L1_single.json
│   ├── L2_objects.json
│   ├── L3_2d_spatial.json
│   ├── L4_occ.json
│   └── ...
├── Spatial457.py               # Hugging Face dataset loader script
├── README.md                   # Documentation

Each JSON file contains a list of VQA examples, where each item includes:

• "image_filename": image file name used in the question
• "question": natural language question
• "answer": boolean, string, or number
• "program": symbolic program (optional)
• "question_index": unique identifier

This modular structure supports scalable multi-task evaluation across levels and reasoning types.

---

🛠️ Dataset Usage

You can load the dataset directly using the Hugging Face 🤗 datasets library:

🔹 Load a specific subtask (e.g., L5_6d_spatial)

from datasets import load_dataset

dataset = load_dataset("RyanWW/Spatial457", name="L5_6d_spatial", split="validation", data_dir=".")

Each example is a dictionary like:

{
  'image': <PIL.Image.Image>,
  'image_filename': 'superCLEVR_new_000001.png',
  'question': 'Is the large red object in front of the yellow car?',
  'answer': 'True',
  'program': [...],
  'question_index': 100001
}

🔹 Other available configurations

[
  "L1_single", "L2_objects", "L3_2d_spatial",
  "L4_occ", "L4_pose", "L5_6d_spatial", "L5_collision"
]

You can swap name="..." in load_dataset(...) to evaluate different spatial reasoning capabilities.

📊 Benchmark

We benchmarked a wide range of state-of-the-art models—including GPT-4o, Gemini, Claude, and several open-source LMMs—across all subsets. The results below have been updated after rerunning the evaluation. While they show minor variance compared to the results in the published paper, the conclusions remain unchanged.

The inference script supports VLMEvalKit and is run by setting the dataset to Spatial457. You can find the detailed inference scripts here.

Spatial457 Evaluation Results

Model	L1_single	L2_objects	L3_2d_spatial	L4_occ	L4_pose	L5_6d_spatial	L5_collision
GPT-4o	72.39	64.54	58.04	48.87	43.62	43.06	44.54
GeminiPro-1.5	69.40	66.73	55.12	51.41	44.50	43.11	44.73
Claude 3.5 Sonnet	61.04	59.20	55.20	40.49	41.38	38.81	46.27
Qwen2-VL-7B-Instruct	62.84	58.90	53.73	26.85	26.83	36.20	34.84

---

📚 Citation

@inproceedings{wang2025spatial457,
  title     = {Spatial457: A Diagnostic Benchmark for 6D Spatial Reasoning of Large Multimodal Models},
  author    = {Wang, Xingrui and Ma, Wufei and Zhang, Tiezheng and de Melo, Celso M and Chen, Jieneng and Yuille, Alan},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2025},
  url       = {https://arxiv.org/abs/2502.08636}
}