oscarqjh/3DSRBench_lmmseval

3DSRBench (lmms-eval compatible)

This is a reformatted version of 3DSRBench for compatibility with lmms-eval.

Dataset Description

3DSRBench is a comprehensive 3D spatial reasoning benchmark that evaluates the 3D spatial reasoning capabilities of Large Multimodal Models (LMMs). It includes 2,100 VQAs on MS-COCO images and 672 on multi-view synthetic images rendered from HSSD.

Subsets

This dataset contains two subsets:

1. 3dsr

• Based on 3dsrbench_v1_vlmevalkit.tsv
• VQA data with question and choices processed with flip augmentation
• 5,157 examples

2. 3dsr_circular

• Based on 3dsrbench_v1_vlmevalkit_circular.tsv
• Augmented with circular evaluation for more robust evaluation
• 11,686 examples
• Compatible with VLMEvalKit data format

Columns

Column	Description
index	Unique identifier for each question
qid	Question ID (only in 3dsr_circular)
question	The question text
A, B, C, D	Multiple choice options
answer	Correct answer (A, B, C, or D)
category	Category of spatial reasoning (height, location, orientation, multi-object)
image	The image for the question
image_source	Source of the image
image_url	Relative path to the COCO image (e.g., /train2017/000000557944.jpg)

Usage

from datasets import load_dataset

# Load the standard subset
dataset = load_dataset("oscarqjh/3DSRBench_lmmseval", "3dsr")

# Load the circular evaluation subset
dataset_circular = load_dataset("oscarqjh/3DSRBench_lmmseval", "3dsr_circular")

# Access an example
example = dataset["test"][0]
print(example["question"])
print(example["image"])  # PIL Image

Additional Files

• coco_images.zip: All MS-COCO images used in 3DSRBench

Citation

@article{ma20243dsrbench,
  title={3DSRBench: A Comprehensive 3D Spatial Reasoning Benchmark},
  author={Ma, Wufei and Chen, Haoyu and Zhang, Guofeng and de Melo, Celso M and Yuille, Alan and Chen, Jieneng},
  journal={arXiv preprint arXiv:2412.07825},
  year={2024}
}

Original Repository

For more information, visit the original 3DSRBench repository and dataset.