[Paper Review] AlignNet-3D: Fast Point Cloud Registration of Partially Observed Objects

Johannes Größ, Aljoša Ošep+1 more|arXiv (Cornell University)|Oct 10, 2019

Robotics and Sensor-Based Localization50 references19 citations

TL;DR

AlignNet-3D proposes a deep learning-based method for fast and accurate 3D point cloud registration of partially observed objects, using a siamese neural network to predict relative transformations between LiDAR scans. It outperforms ICP and global registration methods in accuracy and efficiency, especially under large temporal gaps and varying point densities, while achieving lower velocity estimation error than state-of-the-art baselines.

ABSTRACT

Methods tackling multi-object tracking need to estimate the number of targets in the sensing area as well as to estimate their continuous state. While the majority of existing methods focus on data association, precise state (3D pose) estimation is often only coarsely estimated by approximating targets with centroids or (3D) bounding boxes. However, in automotive scenarios, motion perception of surrounding agents is critical and inaccuracies in the vehicle close-range can have catastrophic consequences. In this work, we focus on precise 3D track state estimation and propose a learning-based approach for object-centric relative motion estimation of partially observed objects. Instead of approximating targets with their centroids, our approach is capable of utilizing noisy 3D point segments of objects to estimate their motion. To that end, we propose a simple, yet effective and efficient network, \method, that learns to align point clouds. Our evaluation on two different datasets demonstrates that our method outperforms computationally expensive, global 3D registration methods while being significantly more efficient. We make our data, code, and models available at https://www.vision.rwth-aachen.de/page/alignnet.

Motivation & Objective

• To address the lack of precise 3D motion estimation in multi-object tracking, especially in close-range autonomous driving scenarios where coarse approximations can lead to catastrophic failures.
• To develop a learning-based approach that leverages full 3D point segments—rather than centroids or bounding boxes—for accurate relative motion estimation between consecutive LiDAR scans.
• To create a robust, efficient, and generalizable method for object-centric 3D registration that handles large temporal gaps, varying point densities, and partial observations.
• To evaluate the method on both synthetic and real-world data, demonstrating generalization to real LiDAR sequences such as KITTI.
• To provide a publicly available codebase and dataset to support future research in 3D point cloud registration and motion estimation.

Proposed method

• A siamese neural network architecture is used to process two point clouds from consecutive time steps, learning a shared representation for each.
• The network first transforms both point clouds into a canonical, amodal pose to reduce sensitivity to initial alignment and viewpoint variation.
• Point cloud embeddings are concatenated and passed through a multi-layer perceptron to predict the relative transformation (translation and rotation) between the two scans.
• The network is trained end-to-end using a loss function that minimizes the difference between predicted and ground-truth transformations on synthetic data.
• The method is optimized for inference speed by leveraging GPU parallelization, achieving ~1.22ms per transformation with a batch size of 32.
• A synthetic dataset is generated from ModelNet40 CAD models, simulating LiDAR point sampling characteristics to train and evaluate the model under controlled conditions.

Experimental results

Research questions

• RQ1 Can a deep learning-based approach achieve higher accuracy than classical ICP and global registration methods in 3D point cloud registration of partially observed objects?
• RQ2 How robust is the proposed method to large temporal gaps and varying point densities in real-world LiDAR sequences?
• RQ3 Can a learnable registration network outperform state-of-the-art motion estimation baselines such as ADH [16] in terms of velocity estimation accuracy?
• RQ4 To what extent does the network generalize from synthetic training data to real-world KITTI tracking sequences?
• RQ5 Is the method efficient enough for real-time application in autonomous vehicle perception systems?

Key findings

• On the KITTI tracking dataset, AlignNet-3D achieves a 14.48% success rate for 2cm, 1° registration accuracy in hard cases (distance < 80m, >10 frame gap), significantly outperforming ICP and FGR variants.
• The method reduces RMSE for translation to 0.48m and rotation to 31.08° on the KITTITrackletsCarsHard set, outperforming global ICP and FGR, which report RMSEs of 38.34m and 70.78° respectively.
• In velocity estimation, AlignNet-3D achieves an RMSE of 0.732m/s at distances < 5m, outperforming ADH 3D (1.002m/s) and the centroid-based Kalman filter (1.604m/s).
• The method is computationally efficient, achieving ~1.22ms per transformation on a GPU with batch size 32, making it suitable for real-time tracking applications.
• Even with a larger temporal gap, AlignNet-3D maintains robust performance, with 11.16% success rate at 20cm, 10° threshold on KITTITrackletsCarsPersonsHard, while ICP and FGR fail completely.
• The network generalizes well to real-world data, achieving performance comparable to computationally expensive global registration methods while being significantly faster.