Point Tree Transformer for Point Cloud Registration

TL;DR

The Point Tree Transformer (PTT) is proposed, a novel transformer-based approach for point cloud registration that efficiently extracts comprehensive local and global features while maintaining linear computational complexity.

Abstract

Point cloud registration is a fundamental task in the fields of computer vision and robotics. Recent developments in transformer-based methods have demonstrated enhanced performance in this domain. However, the standard attention mechanism utilized in these methods often integrates many low-relevance points, thereby struggling to prioritize its attention weights on sparse yet meaningful points. This inefficiency leads to limited local structure modeling capabilities and quadratic computational complexity. To overcome these limitations, we propose the Point Tree Transformer (PTT), a novel transformer-based approach for point cloud registration that efficiently extracts comprehensive local and global features while maintaining linear computational complexity. The PTT constructs hierarchical feature trees from point clouds in a coarse-to-dense manner, and introduces a novel Point Tree Attention (PTA) mechanism, which follows the tree structure to facilitate the progressive convergence of attended regions towards salient points. Specifically, each tree layer selectively identifies a subset of key points with the highest attention scores. Subsequent layers focus attention on areas of significant relevance, derived from the child points of the selected point set. The feature extraction process additionally incorporates coarse point features that capture high-level semantic information, thus facilitating local structure modeling and the progressive integration of multiscale information. Consequently, PTA empowers the model to concentrate on crucial local structures and derive detailed local information while maintaining linear computational complexity. Extensive experiments conducted on the 3DMatch, ModelNet40, and KITTI datasets demonstrate that our method achieves superior performance over the state-of-the-art methods.

Figures (11)

• Figure 1: Explanation of the Point Tree Attention (PTA) and a comparison to the attention mechanisms in the Standard Transformer (ST) vaswani2017attention and Point Transformer (PT) zhao2021point, with visualization of attention weights for the point marked with a green dot. (a) In our method, feature trees are built, and then PTA is used to hierarchically incorporate coarse features and restrict the attended regions of the next layer to the child points of the top $\mathcal{S}$ keys with the highest attention scores, skipping the shaded regions, where the locations of the top $\mathcal{S}$ keys are highlighted in the same color as the query. Therefore, (b) PTA can focus on critical local structures and adaptively attend to relevant regions. In comparison, ST considers many low-relevance points and struggles to capture the local structures, whereas PT simply sets attention regions to predefined templates, overlooking information from other relevant regions.
• Figure 2: Network architecture of the PTT. First, the KPConv extracts features for a sparse set of points. Subsequently, the tree transformer encoder builds feature trees and iteratively extracts features containing local and global information. Then, the decoder predicts the corresponding point clouds and overlap scores. Finally, a transformation is computed to align the point clouds.
• Figure 3: Illustration of tree construction. The point cloud in the densest layer utilizes the input point cloud and the sub-dense point cloud is obtained by voxelizing the densest point cloud. Then, the following layers group $\mathbb{N}$ voxels into one voxel to obtain the point cloud.
• Figure 4: (a) Illustration of the PTA module, where $\uparrow$ indicates the selected top $\mathcal{S}$ key points. (b) PTA incorporates spatially coarse features into the corresponding child points to guide the feature extraction process. (c) Additionally, PTA hierarchically selects the top $\mathcal{S}$ key points with the highest attention scores. In each layer beyond the first, attention is evaluated only in the specified regions, which correspond to the child points of the $\mathcal{S}$ selected coarse key points. These specified regions are highlighted in the same color as the query.
• Figure 5: Qualitative registration results obtained on (a, b) 3DMatch, (c, d) 3DLoMatch, (e) ModelNet40, and (f) KITTI.