SpaceDG-Bench: 1,102 Questions for MLLM Spatial Intelligence Under Visual Degradation

SpaceDG-Bench is a human-verified benchmark containing 1,102 questions designed to evaluate the spatial intelligence of Multimodal Large Language Models (MLLMs) under visual degradation. The dataset spans 11 reasoning categories and 9 visual degradation types, yielding over 10,000 visual question answering (VQA) instances. It was created by author xlzhou126 and last updated on May 24, 2026.

Use Cases

• Benchmarking MLLM spatial reasoning capabilities based on the 11 reasoning categories
• Testing model robustness to visual artifacts based on the 9 degradation types like motion blur and low light
• Evaluating visual question answering performance under adverse conditions based on the over 10,000 VQA instances

Strengths

• Contains 1,102 questions, providing a substantial test set
• Covers 9 distinct visual degradation types, including motion blur and adverse weather
• Yields over 10,000 VQA instances from the question and degradation combinations
• Is described as human-verified, suggesting a level of quality control

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count is unknown, which may limit suitability assessment
• Description metadata is limited; actual data quality requires manual inspection after download

Provenance

Source

xlzhou126 on Hugging Face, part of the SpaceDG project

Collection Method

Likely constructed for benchmark evaluation; described as human-verified

Freshness

Last updated 2026-05-24 16:01:15; freshness should be verified