EADReg: Probabilistic Correspondence Generation with Efficient Autoregressive Diffusion Model for Outdoor Point Cloud Registration

TL;DR

A novel framework named EADReg is proposed for efficient and robust registration of LiDAR point clouds based on autoregressive diffusion models that achieves runtime comparable to convolutional-based methods.

Abstract

Diffusion models have shown the great potential in the point cloud registration (PCR) task, especially for enhancing the robustness to challenging cases. However, existing diffusion-based PCR methods primarily focus on instance-level scenarios and struggle with outdoor LiDAR points, where the sparsity, irregularity, and huge point scale inherent in LiDAR points pose challenges to establishing dense global point-to-point correspondences. To address this issue, we propose a novel framework named EADReg for efficient and robust registration of LiDAR point clouds based on autoregressive diffusion models. EADReg follows a coarse-to-fine registration paradigm. In the coarse stage, we employ a Bi-directional Gaussian Mixture Model (BGMM) to reject outlier points and obtain purified point cloud pairs. BGMM establishes correspondences between the Gaussian Mixture Models (GMMs) from the source and target frames, enabling reliable coarse registration based on filtered features and geometric information. In the fine stage, we treat diffusion-based PCR as an autoregressive process to generate robust point correspondences, which are then iteratively refined on upper layers. Despite common criticisms of diffusion-based methods regarding inference speed, EADReg achieves runtime comparable to convolutional-based methods. Extensive experiments on the KITTI and NuScenes benchmark datasets highlight the state-of-the-art performance of our proposed method. Codes will be released upon publication.

Figures (4)

• Figure 1: Comparison with previous Diffusion PCR methods. Previous diffusion-based PCR methods directly generate global dense point-to-point (P2P) correspondences $C_{t}\in\mathbb{R}^{N^{\mathcal{S}}\times N^{\mathcal{T}}}$, resulting in high training costs. In contrast, our proposed EADReg predicts correspondences within the top-K nearest neighbors of the source points $\hat{C}\in\mathbb{R}^{N^{\mathcal{S}}\times K}$, leveraging reliable coarse registration. Besides, We integrate an autoregressive framework into the reverse process to better suit PCR tasks.
• Figure 2: The detector-descriptor backbone hierarchically downsamples the input point cloud pairs and extracts corresponding features. Next, the BGMM Outlier Removal module purifies the input points and leverages the predictor network $\mathcal{F}_{c}$ along with SVD to perform coarse registration. Finally, our proposed AD inference framework autoregressively generates regional correspondences and predicts the final registration result.
• Figure 3: Qualitative visualization of registration performance. From left to right, we compare our proposed method with HRegNet using three samples from the KITTI, NuScenes, and Apollo-Southbay datasets, respectively. Specifically, we select only the correspondences with confidence weights $\hat{w}$ greater than 0.001.
• Figure 4: We select four samples from the KITTI dataset to illustrate the outlier removal process using the BGMM module. The first row contains the original point clouds, and the second row shows the purified version. The outliers are framed within boxes.