Learning-based registration for large-scale 3D point clouds has been shown to improve robustness and accuracy compared to classical methods and can be trained without supervision for locally rigid problems. However, for tasks with highly deformable structures, such as alignment of pulmonary vascular trees for medical diagnostics, previous approaches of self-supervision with regularisation and point distance losses have failed to succeed, leading to the need for complex synthetic augmentation strategies to obtain reliably strong supervision. In this work, we introduce a novel Differentiable Volumetric Rasterisation of point Clouds (DiVRoC) that overcomes those limitations and offers a highly efficient and accurate loss for large-scale deformable 3D registration. DiVRoC drastically reduces the computational complexity for measuring point cloud distances for high-resolution data with over 100k 3D points and can also be employed to extrapolate and regularise sparse motion fields, as loss in a self-training setting and as objective function in instance optimisation. DiVRoC can be successfully embedded into geometric registration networks, including PointPWC-Net and other graph CNNs. Our approach yields new state-of-the-art accuracy on the challenging PVT dataset in three different settings without training with manual ground truth: 1) unsupervised metric-based learning 2) self-supervised learning with pseudo labels generated by self-training and 3) optimisation based alignment without learning. https://github.com/mattiaspaul/ChasingClouds

A popular benchmark for intra-patient lung registration is provided by the DIR-LAB COPDgene dataset consisting of large-motion in- and expiratory breath-hold CT pairs. This dataset alone, however, does not provide enough samples to properly train state-of-the-art deep learning methods. Other public datasets often also provide only small sample sizes or include primarily small motions between scans that do not translate well to larger deformations. For point-based geometric registration, the PVT1010 dataset provides a large number of vessel point clouds without any correspondences and a labeled test set corresponding to the COPDgene cases. However, the absence of correspondences for supervision complicates training, and a fair comparison with image-based algorithms is infeasible, since CT scans for the training data are not publicly available.We here provide a combined benchmark for image- and point-based registration approaches. We curated a total of 248 public multi-centric in- and expiratory lung CT scans from 124 patients, which show large motion between scans, processed them to ensure sufficient homogeneity between the data and generated vessel point clouds that are well distributed even deeper inside the lungs. For supervised training, we provide vein and artery segmentations of the vessels and multiple thousand image-derived keypoint correspondences for each pair. For validation, we provide multiple scan pairs with manual landmark annotations. Finally, as first baselines on our new benchmark, we evaluate several image and point cloud registration methods on the dataset.

Flow estimation on 3D point clouds is a challenging problem in the field of computer vision, which has great significance in many areas, such as autonomous driving and human interaction applications. Within the last years, the field of motion analysis has made great progress. The evaluation of the existing approaches mostly focuses on scenarios where objects are affected by rigid transformations. However, in many application areas such as gesture recognition or pose tracking, the detection of shape changes is essential and breaking them down to local rigid transformations is accompanied by loss of information. One component of our contributions is that we specifically prepared existing datasets for scene flow estimation on deformable objects. Additionally, we benchmark existing methods and analyze their behavior on various subtasks. The results show that already close to 80% of correct correspondences can be found on synthetic hand data, while only around 50% are found on real hand data. Our experimental validation and analysis help to build an understanding of new possibilities in broader areas. Furthermore, they should help to inspire possible further research directions.

Hand gesture recognition can benefit from directly processing 3D point cloud sequences, which carry rich geometric information and enable the learning of expressive spatio-temporal features. However, currently employed single-stream models cannot sufficiently capture multi-scale features that include both fine-grained local posture variations and global hand movements. We therefore propose a novel dual-stream model, which decouples the learning of local and global features. These are eventually fused in an LSTM for temporal modelling. To induce the global and local stream to capture complementary position and posture features, we propose the use of different 3D learning architectures in both streams. Specifically, state-of-the-art point cloud networks excel at capturing fine posture variations from raw point clouds in the local stream. To track hand movements in the global stream, we combine an encoding with residual basis point sets and a fully-connected DenseNet. We evaluate the method on the Shrec'17 and DHG dataset and report state-of-the-art results at a reduced computational cost. Source code is available at https://github.com/multimodallearning/hand-gesture-posture-position.