New Dataset & Benchmark: SPAR-7M and SPAR-Bench for Spatial Perception and Reasoning with VLMs

[hutchinsonian]

Hi,

Appreciate your outstanding effort in maintaining this great resource for the MLLM community.

We would like to recommend our recent work “From Flatland to Space: Teaching Vision-Language Models to Perceive and Reason in 3D”, which focuses on enabling VLMs to perform 3D spatial reasoning using only 2D observations. We introduce:

• SPAR-7M: a large-scale dataset with 7M QA pairs across 33 spatial tasks.
• SPAR-Bench: a comprehensive benchmark covering both low-level perception and high-level 3D reasoning in single-view and multi-view settings.

We hope our work can be considered for inclusion in the Datasets & Benchmarks or Multimodal Spatial Understanding section.

• 📄 Paper: https://arxiv.org/abs/2503.22976
• 🔗 GitHub: https://github.com/fudan-zvg/SPAR
• 🌐 Project Page: https://fudan-zvg.github.io/spar

Thank you!