SpaceDG: Benchmarking Spatial Intelligence under Visual Degradation

2026-05-21

AI Summary

This work addresses the critical gap in existing spatial reasoning benchmarks, which neglect realistic visual degradations and thus fail to assess the robustness of multimodal large language models (MLLMs) under non-ideal conditions. To bridge this gap, we introduce SpaceDG, the first large-scale dataset for spatial understanding under visual degradation, leveraging a physics-driven degradation synthesis engine combined with 3D Gaussian splatting rendering to generate over one million question-answer pairs spanning nine degradation types. We further establish SpaceDG-Bench, a human-verified evaluation benchmark, and propose a degradation-aware training paradigm—the first to incorporate photorealistic visual degradations into spatial intelligence assessment. Experiments reveal that 25 state-of-the-art MLLMs suffer significant performance drops under degradation, whereas models fine-tuned on SpaceDG not only surpass human performance in degraded conditions but also maintain full accuracy on clean images.

Abstract

Multimodal Large Language Models (MLLMs) have made rapid progress in spatial intelligence, yet existing spatial reasoning benchmarks largely assume pristine visual inputs and overlook the degradations that commonly occur in real-world deployment, such as motion blur, low light, adverse weather, lens distortion, and compression artifacts. This raises a fundamental question: how robust is the spatial intelligence of current MLLMs when visual observations are imperfect? To answer this question, we introduce SpaceDG, the first large-scale dataset for degradation-aware spatial understanding. It is constructed with a physically grounded degradation synthesis engine that embeds degradation formation process into 3D Gaussian Splatting (3DGS) rendering, enabling realistic simulation of nine degradation types. The resulting dataset contains approximately 1M QA pairs from nearly 1,000 indoor scenes. We further introduce SpaceDG-Bench, an human-verified benchmark with 1,102 questions spanning 11 reasoning categories and 9 visual degradation types, yielding over 10K VQA instances. Evaluating 25 open- and closed-source MLLMs reveals that visual degradations consistently and substantially impair spatial reasoning, exposing a critical robustness gap. Finally, we show that finetuning on SpaceDG markedly improves degradation robustness and can even surpass human performance under degraded conditions without any performance drop on clean images, highlighting the promise of degradation-aware training for robust spatial intelligence.

Problem

• spatial intelligence
• visual degradation
• multimodal large language models
• robustness
• benchmarking

Innovation

• visual degradation
• spatial intelligence
• 3D Gaussian Splatting
• multimodal large language models
• robustness benchmark