Scholarlink

<p>Indoor 3D object detection often relies on superpoints as intermediate query units. However, superpoints generated purely from local geometry are easily contaminated by background structures under occlusion and clutter, leading to unstable predictions. We propose a mask-aware tri-modal framework that improves the quality of superpoint representations. First, a scene-level structural context is retrieved from a pretrained PointSAM encoder to enhance object-centric evidence. Second, the Mask-Aware Feature Refinement Module (MFR) predicts a soft mask weight to suppress unreliable superpoints. Finally, visual-language priors derived from bird’s-eye-view (BEV) maps and scene text are fused with geometric tokens to reduce semantic ambiguity. Experiments on ScanNet and S3DIS show that the proposed method achieves 78.52 mAP@0.25 and 66.70 mAP@0.5 on ScanNet, while keeping the downstream detector architecture unchanged. For the video demonstration, please refer to <a href="https://youtu.be/11uDANt72so">https://youtu.be/11uDANt72so</a>. Code will be available upon the paper’s acceptance.</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

Mask-aware tri-modal learning for indoor 3D object detection

  • 期刊论文

  • 发表日期：

    2026年7月17日

查看全文
  • Feng Zhou,
  • Hui Wang,
  • Kaida Ning,
  • Junjun Pan,
  • Jin Li,
  • Ju Dai

摘要

Indoor 3D object detection often relies on superpoints as intermediate query units. However, superpoints generated purely from local geometry are easily contaminated by background structures under occlusion and clutter, leading to unstable predictions. We propose a mask-aware tri-modal framework that improves the quality of superpoint representations. First, a scene-level structural context is retrieved from a pretrained PointSAM encoder to enhance object-centric evidence. Second, the Mask-Aware Feature Refinement Module (MFR) predicts a soft mask weight to suppress unreliable superpoints. Finally, visual-language priors derived from bird’s-eye-view (BEV) maps and scene text are fused with geometric tokens to reduce semantic ambiguity. Experiments on ScanNet and S3DIS show that the proposed method achieves 78.52 mAP@0.25 and 66.70 mAP@0.5 on ScanNet, while keeping the downstream detector architecture unchanged. For the video demonstration, please refer to https://youtu.be/11uDANt72so. Code will be available upon the paper’s acceptance.