PREDATOR: Registration of 3D Point Clouds with Low Overlap

By Shengyu Huang, Zan Gojcic, Mikhail (Misha) Usvyatsov, Andreas Wieser, Konrad Schindler. Computer Vision and Pattern Recognition, 2020

728 citations recorded by the source corpus.

We introduce PREDATOR, a model for pairwise point-cloud registration with deep attention to the overlap region. Different from previous work, our model is specifically designed to handle (also) point-cloud pairs with low overlap. Its key novelty is an overlap-attention block for early information exchange between the latent encodings of the two point clouds. In this way the subsequent decoding of the latent representations into per-point features is conditioned on the respective other point cloud, and thus can predict which points are not only salient, but also lie in the overlap region between the two point clouds. The ability to focus on points that are relevant for matching greatly improves performance: PREDATOR raises the rate of successful registrations by more than 20% in the low-overlap scenario, and also sets a new state of the art for the 3DMatch benchmark with 89% registration recall. [Code release]

Subfields: Aerospace Engineering, Computational Mechanics, Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/06d0515d11386bb1572187f81af5abda655c941e

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Aerospace Engineering

1. Fu Li
2. Benjamin Busam
3. Mahdi Saleh
4. Zixin Luo
5. Lei Zhou
6. Xuyang Bai
7. Hongbo Fu
8. Chiew-Lan Tai

In Computational Mechanics

1. Fu Li
2. Benjamin Busam
3. Mahdi Saleh
4. Zixin Luo
5. Xuyang Bai
6. Lei Zhou
7. Chiew-Lan Tai
8. Hongbo Fu

In Computer Vision and Pattern Recognition

1. Slobodan Ilic
2. Jan Kautz
3. Yuxing Peng
4. Zixin Luo
5. Bowen Wen
6. Lei Li
7. Zheng Qin
8. Lei Zhou

Citations in the corpus

• CoFiNet: Reliable Coarse-to-fine Correspondences for Robust Point Cloud Registration (2021, methodology): “Compared to [12], we only use around two-third and one-fourth of parameters on indoor and outdoor benchmarks, respectively.”
• PointDSC: Robust Point Cloud Registration using Deep Spatial Consistency (2021, methodology): “Recently, learningbased algorithms have been proposed to replace the individual components in the classical registration pipeline, including keypoint detection [4, 38, 32] and feature description [21, 22, 23, 51, 4, 18, 27, 30, 2].”
• Geometric Transformer for Fast and Robust Point Cloud Registration (2022, methodology): “We use the training data preprocessed by [15] and evaluate on both 3DMatch and 3DLoMatch [15] protocols.”
• FoundationPose: Unified 6D Pose Estimation and Tracking of Novel Objects (2023, methodology): “For the model-free setup, a number of reference images capturing the novel object PREDATOR[27] LoFTR[ are selected from the training split of the datasets, equipped with the ground-truth annotation of the object pose, following [21].”