Scholarlink

<p>3D indoor room generation plays a critical role in embodied artificial intelligence, which explores downstream tasks such as environment conception, vision grounding, visual-language navigation, and 3D visual question answering. Although previous methods generate vanilla 3D indoor scenes with reasonable layouts, they are usually limited by the scene-level diversity and insufficient furnish decoration of existing datasets to generate realistic rooms with diverse objects of furniture. Overcoming the labor-intensive data collection procedure, we propose a large-scale 3D room dataset, <b>FurniScene</b>, with intricate <b>Furni</b>shing <b>Scene</b>s from interior design professionals to serve the community. Specifically, the FurniScene consists of 111,698 rooms and 39,691 unique furniture CAD models with 89 different categories, covering things from large beds to small teacups on the coffee table. Different from existing datasets focusing on large furniture, FurniScene pays more attention to detailed furniture decorations, such as teawares, vases, and lamps on the table, which derives a fine-grained indoor scene layout generation task. To address this issue, we also introduce a baseline method named the <b>f</b>urniture-list and <b>l</b>ayout <b>g</b>eneration <b>m</b>odel (FLGM) for fine-grained indoor layout generation. Moreover, we also conduct an evaluation benchmark for various indoor scene generation methods based on the proposed FurniScene. Quantitative and qualitative evaluations demonstrate the capability of our method to generate highly realistic indoor scenes. Our dataset and code will be publicly available soon.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

FurniScene: A Large-scale 3D Room Dataset with Intricate Furnishing Scenes

  • 期刊论文

  • 发表日期：

    2026年2月21日

查看全文
  • Yuxi Wang,
  • Junran Peng,
  • Genghao Zhang,
  • Chuanchen Luo,
  • Shibiao Xu,
  • Man Zhang,
  • Zhaoxiang Zhang

摘要

3D indoor room generation plays a critical role in embodied artificial intelligence, which explores downstream tasks such as environment conception, vision grounding, visual-language navigation, and 3D visual question answering. Although previous methods generate vanilla 3D indoor scenes with reasonable layouts, they are usually limited by the scene-level diversity and insufficient furnish decoration of existing datasets to generate realistic rooms with diverse objects of furniture. Overcoming the labor-intensive data collection procedure, we propose a large-scale 3D room dataset, FurniScene, with intricate Furnishing Scenes from interior design professionals to serve the community. Specifically, the FurniScene consists of 111,698 rooms and 39,691 unique furniture CAD models with 89 different categories, covering things from large beds to small teacups on the coffee table. Different from existing datasets focusing on large furniture, FurniScene pays more attention to detailed furniture decorations, such as teawares, vases, and lamps on the table, which derives a fine-grained indoor scene layout generation task. To address this issue, we also introduce a baseline method named the furniture-list and layout generation model (FLGM) for fine-grained indoor layout generation. Moreover, we also conduct an evaluation benchmark for various indoor scene generation methods based on the proposed FurniScene. Quantitative and qualitative evaluations demonstrate the capability of our method to generate highly realistic indoor scenes. Our dataset and code will be publicly available soon.