PPF-FoldNet: Unsupervised Learning of Rotation Invariant 3D Local Descriptors

By Haowen Deng, Tolga Birdal, Slobodan Ilic. European Conference on Computer Vision, 2018

435 citations recorded by the source corpus.

We present PPF-FoldNet for unsupervised learning of 3D local descriptors on pure point cloud geometry. Based on the folding-based auto-encoding of well known point pair features, PPF-FoldNet offers many desirable properties: it necessitates neither supervision, nor a sensitive local reference frame, benefits from point-set sparsity, is end-to-end, fast, and can extract powerful rotation invariant descriptors. Thanks to a novel feature visualization, its evolution can be monitored to provide interpretable insights. Our extensive experiments demonstrate that despite having six degree-of-freedom invariance and lack of training labels, our network achieves state of the art results in standard benchmark datasets and outperforms its competitors when rotations and varying point densities are present. PPF-FoldNet achieves 9% higher recall on standard benchmarks, 23% higher recall when rotations are introduced into the same datasets and finally, a margin of >35% is attained when point density is significantly decreased.

Subfields: Aerospace Engineering, Computational Mechanics, Geology

View Paper: https://www.semanticscholar.org/paper/9ccef6c69e5c16ac2ce9da4f1411415e402dd59c

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

In Computational Mechanics

In Geology

Citations in the corpus

Rotation Invariant Convolutions for 3D Point Clouds Deep Learning (2019, background): “In addition, there are a few previous works about learning local descriptors from point clouds for feature matching [43, 14, 7], some of which [7] can be rotation invariant.”
PointContrast: Unsupervised Pre-training for 3D Point Cloud Understanding (2020, background): “In the past few years, the ﬁeld of 3D deep learning has witnessed much progress with an ever-increasing number of 3D representation learning schemes [1, 16, 74, 21, 36, 67, 22, 15, 81, 12, 9].”
Deep Global Registration (2020, methodology): “We follow the train/test split and the standard procedure to generate pairs with at least 30% overlap for training and testing [12, 11, 9].”
A comprehensive survey on point cloud registration (2021, methodology): “To solve this issue, PPF-FoldNet [18] proposes an unsupervised method to remove the annotation requirement constraint.”
CoFiNet: Reliable Coarse-to-fine Correspondences for Robust Point Cloud Registration (2021, methodology): “PPF-FoldNet [8] leverages only PPF, which is naturally rotation-invariant, as its input and further incorporates a FoldingNet [29] architecture to enable the unsupervised training of rotation-invariant descriptors.”
Global-Local Bidirectional Reasoning for Unsupervised Representation Learning of 3D Point Clouds (2020, methodology): “Self-Reconstruction: Self-reconstruction, or point autoencoding, is a widely used technique for unsupervised point cloud representation learning [1,8,14,26,47,55,56].”
SpinNet: Learning a General Surface Descriptor for 3D Point Cloud Registration (2020, methodology): “By comparison, our SpinNet achieves an overall FMR score of 79.6%, demonstrating the superior generalization across novel scenarios.”
D3Feat: Joint Learning of Dense Detection and Description of 3D Local Features (2020, background): “[5, 4] build their network upon PointNet to directly consume unordered point sets.”
3D Point Capsule Networks (2018, methodology): “PPF local encoding of PPF-FoldNet is rotation-invariant.”
PREDATOR: Registration of 3D Point Clouds with Low Overlap (2020, background): “PPFNet [10], and its self-supervised version PPFFoldNet [9], combine point pair features with a PointNet [28] architecture to extract descriptors that are aware of the global context.”
The Perfect Match: 3D Point Cloud Matching With Smoothed Densities (2018, methodology): “Since achieving the rotation invariance in practice is non-trivial, several recent works [49, 5, 45] choose to ignore it and thus do not generalize to rigidly transformed scenes [4].”
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].”
DiffusionNet: Discretization Agnostic Learning on Surfaces (2020, background): “Moreover, recent efforts have also been made to incorporate invariance and equivariance of the networks with respect to various geometric transformations, e.g. , [Deng et al. 2018; Hansen et al. 2018; Li et al. 2021; Poulenard et al. 2019; Zhang et al. 2019; Zhao et al. 2020].”
Deep Closest Point: Learning Representations for Point Cloud Registration (2019, methodology): “Recently, descriptor learning methods have brought significant progress in point cloud registration: 3DMatch [59] proposes learning a local volumetric patch descriptor to establish correspondences; 3DFeatNet [56] takes similar approach to point cloud representation for local regions; PPFFoldNet [9] ”
Patch-Based Progressive 3D Point Set Upsampling (2018, background): “A few recent works have exploited the power of multiscale representation [12, 24, 28, 37, 49] and skipconnection [8,43] in 3D learning.”
Fully Convolutional Geometric Features (2019, methodology): “As noted in several works [1, 7, 6, 17], recall is more important than precision since it is possible to improve precision with better pruning.”
DeepVCP: An End-to-End Deep Neural Network for Point Cloud Registration (2019, background): “We are also considering to try rotation invariant 3D descriptors [6, 16, 23] in the future.”