[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-81014":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":10,"language":11,"languages":10,"totalLinesOfCode":10,"stars":12,"forks":13,"watchers":14,"openIssues":15,"contributorsCount":13,"subscribersCount":13,"size":13,"stars1d":13,"stars7d":13,"stars30d":15,"stars90d":13,"forks30d":13,"starsTrendScore":13,"compositeScore":13,"rankGlobal":10,"rankLanguage":10,"license":16,"archived":17,"fork":17,"defaultBranch":18,"hasWiki":17,"hasPages":19,"topics":20,"createdAt":10,"pushedAt":10,"updatedAt":26,"readmeContent":27,"aiSummary":28,"trendingCount":13,"starSnapshotCount":13,"syncStatus":29,"lastSyncTime":30,"discoverSource":31},81014,"SpaceDG","Visionary-Laboratory\u002FSpaceDG","Visionary-Laboratory","SpaceDG: Benchmarking Spatial Intelligence under Visual Degradation","https:\u002F\u002Fvisionary-laboratory.github.io\u002FSpaceDG\u002F",null,"Python",30,0,29,1,"Apache License 2.0",false,"main",true,[21,22,23,24,25],"low-level-vision","mllm","mllm-evaluation","spatial-intelligence","vlm","2026-06-12 02:04:09","# \u003Cimg src=\"assets\u002Flogo.png\" width=\"45\" \u002F> SpaceDG: Benchmarking Spatial Intelligence under Visual Degradation\n\n[🌐 Homepage](https:\u002F\u002Fvisionary-laboratory.github.io\u002FSpaceDG\u002F) | [🤗 Benchmark](https:\u002F\u002Fhuggingface.co\u002Fdatasets\u002Fxlzhou126\u002FSpaceDG-Bench) | [📖 arXiv](https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.22536)\n\n![SpaceDG teaser figure](assets\u002Fteaser.png)\n\n## Overview\n\nMultimodal Large Language Models (MLLMs) have improved spatial reasoning, yet most benchmarks assume pristine images and ignore real degradations such as motion blur, low light, adverse weather, lens distortion, and compression. This raises a fundamental question: How robust is spatial intelligence when observations are imperfect? To address this question, we introduce **SpaceDG**, the first large-scale dataset for degradation-aware spatial understanding: a physically grounded synthesis pipeline embeds nine degradation types into 3D Gaussian Splatting rendering, yielding roughly 1M QA pairs across nearly 1,000 indoor scenes. We further release **SpaceDG-Bench**, a human-verified benchmark with 1,102 questions spanning 11 reasoning categories and 9 degradation types (10K+ VQA instances). We conduct a comprehensive evaluation of 25 open- and closed-source models. Our analysis identifies four key findings:\n\n- **First**, visual degradations consistently impair spatial reasoning across all evaluated MLLMs, highlighting the need for degradation-aware spatial evaluation.\n- **Second**, humans also suffer clear performance drops under degraded conditions. This suggests that the design of MLLMs should not simply imitate human perception, but should learn degradation-aware spatial knowledge to better handle diverse real-world visual inputs.\n- **Third**, degradation-based supervised fine-tuning yields substantial improvements on both clean and degraded inputs, indicating that exposure to physically grounded degradations can enhance robust spatial understanding.\n- **Finally**, visual degradations affect fine-grained object-level perception (such as object counting) more strongly than certain geometric reasoning tasks (such as camera-centric translation), revealing that detailed visual grounding is particularly sensitive to degraded visual evidence.\n\n## Quick Start (EASI Evaluation)\n\n### 1) Environment Setup\n\nUse the EASI setup script to prepare the runtime environment with uv.\n\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002FVisionary-Laboratory\u002FSpaceDG.git\ncd SpaceDG\u002FEASI\nbash scripts\u002Fsetup.sh\n```\n\n### 2) Prepare Data\n\nGet **SpaceDG-Bench** from Hugging Face (file layout and notes are on the dataset card):\n\n- [`xlzhou126\u002FSpaceDG-Bench`](https:\u002F\u002Fhuggingface.co\u002Fdatasets\u002Fxlzhou126\u002FSpaceDG-Bench)\n\n**Default:** the first time you evaluate with `--data spacedg_bench`, the VLMEvalKit dataset loader downloads `spacedg_bench.tsv` and the parquet shards, runs in-repo image extraction (`prepare_data` inside `spacedg_bench.py`), and caches assets under **`~\u002FLMUData`**. You do **not** need a separate `prepare_data.py` script.\n\n**Offline \u002F pre-downloaded tree:** if you already have a directory containing `spacedg_bench.tsv` and the image files so every `image_path` in the TSV resolves, set:\n\n```bash\nexport SPACEDG_BENCH_ROOT=\u002Fpath\u002Fto\u002FSpaceDG_Bench\n```\n\nThat skips automatic downloads for this benchmark. Otherwise, follow the usual VLMEvalKit \u002F EASI environment setup.\n\n### 3) Evaluation with VLMEvalKit\n\nWe provide an example launcher script:\n\n- `EASI\u002FVLMEvalKit\u002Fscripts\u002Frun_spacedg_bench.sh`\n\nOr run `torchrun` directly from the VLMEvalKit root:\n\n```bash\ncd \u003CPATH_TO_THIS_REPO>\u002FSpaceDG\u002FEASI\u002FVLMEvalKit\n\nCUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 torchrun run.py \\\n  --model InternVL3_5-8B \\\n  --data spacedg_bench \\\n  --mode all \\\n  --work-dir ..\u002Foutputs_spacedg \\\n  --reuse\n```\n\n## TODO\n\n- [ ] Release full SpaceDG dataset.\n\n- [x] Release SpaceDG-Bench and evaluation code.\n\n- [x] Release the full paper and the project page of SpaceDG.\n\n\u003C!-- ## Citation\n\nIf you find this work useful, please consider citing:\n\n```bibtex\n@article{zhou2026spacedg,\n  title={SpaceDG: Benchmarking Spatial Intelligence under Visual Degradation},\n  author={Zhou, Xiaolong and Liu, Yifei and Gong, Ziyang and Li, Jiarui and Zhao, Qiyue and Niu, Muyao and Ma, Le and Yang, Xue and Zhang, Hongjie and Zhong, Zhihang},\n  journal={arXiv preprint arXiv:2605.22536},\n  year={2026}\n}\n``` -->\n","SpaceDG是一个用于评估在视觉退化条件下空间智能表现的基准项目。它通过物理基础的合成管道将九种退化类型嵌入到3D高斯点云渲染中，生成了约100万个问答对，覆盖近1000个室内场景。此外，该项目还提供了一个经过人工验证的基准测试集SpaceDG-Bench，包含11个推理类别和9种退化类型的1,102个问题（超过10,000个VQA实例）。SpaceDG适用于需要评估多模态大语言模型在面对诸如运动模糊、低光照等真实世界视觉退化情况下的性能的场景。通过分析发现，视觉退化不仅影响机器模型的空间推理能力，也降低了人类的表现，强调了设计能够处理多种视觉输入的鲁棒性模型的重要性。",2,"2026-06-11 04:03:10","CREATED_QUERY"]