Scholarlink

<p>This work presents CAT++, an enhanced Context-Aware Transformer (CAT), to address the need for more effective 3D point cloud annotation. CAT++ is an end-to-end 3D automatic annotator that lifts 2D weak annotations into 3D, thereby minimizing human annotation burdens. Specifically, it employs a hierarchical-interleaved encoder (HIE) to enhance its encoder-decoder architecture. The HIE alternates between intra- and inter-object Transformer blocks, integrating local and global object features by performing self-attentions along the sequence and batch dimensions. This process continuously combines local and global object features and creates a 3D representation that links interactions between objects at different scales, leading to a more comprehensive understanding of the scene. Moreover, CAT++ features an attention-conditioned implicit neural representation (INR) network, which has been specially engineered to enable more precise and efficient continuous modeling of 3D object surfaces. This improves CAT++’s multi-task learning and its ability to handle hard samples. Results on KITTI and nuScenes datasets show CAT++’s state-of-the-art performance, outperforming other methods by up to 2% <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(AP_{3D}\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>A</mi> <msub> <mi>P</mi> <mrow> <mn>3</mn> <mi>D</mi> </mrow> </msub> </mrow> </math></EquationSource> </InlineEquation> on the KITTI <i>test</i> split. Furthermore, CAT++ significantly reduces human annotation effort, by factors of 6.4 and 50 on the KITTI and nuScenes datasets for the Car category, respectively.</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

CAT++: Enhancing 3D Annotations with Hierarchical-Interleaved Encoding and Attention-Conditioned Implicit Representation

  • 期刊论文

  • 发表日期：

    2026年1月20日

查看全文
  • Xiaoyan Qian,
  • Chang Liu,
  • Xiaojuan Qi,
  • Siewchong Tan,
  • Edmund Y. Lam,
  • Ngai Wong

摘要

This work presents CAT++, an enhanced Context-Aware Transformer (CAT), to address the need for more effective 3D point cloud annotation. CAT++ is an end-to-end 3D automatic annotator that lifts 2D weak annotations into 3D, thereby minimizing human annotation burdens. Specifically, it employs a hierarchical-interleaved encoder (HIE) to enhance its encoder-decoder architecture. The HIE alternates between intra- and inter-object Transformer blocks, integrating local and global object features by performing self-attentions along the sequence and batch dimensions. This process continuously combines local and global object features and creates a 3D representation that links interactions between objects at different scales, leading to a more comprehensive understanding of the scene. Moreover, CAT++ features an attention-conditioned implicit neural representation (INR) network, which has been specially engineered to enable more precise and efficient continuous modeling of 3D object surfaces. This improves CAT++’s multi-task learning and its ability to handle hard samples. Results on KITTI and nuScenes datasets show CAT++’s state-of-the-art performance, outperforming other methods by up to 2% \(AP_{3D}\) A P 3 D on the KITTI test split. Furthermore, CAT++ significantly reduces human annotation effort, by factors of 6.4 and 50 on the KITTI and nuScenes datasets for the Car category, respectively.