No abstract

In this paper, we propose a novel approach, 3D-RecGAN++, which reconstructs the complete 3D structure of a given object from a single arbitrary depth view using generative adversarial networks. Unlike existing work which typically requires multiple views of the same object or class labels to recover the full 3D geometry, the proposed 3D-RecGAN++ only takes the voxel grid representation of a depth view of the object as input, and is able to generate the complete 3D occupancy grid with a high resolution of $256^3$ by recovering the occluded/missing regions. The key idea is to combine the generative capabilities of 3D encoder-decoder and the conditional adversarial networks framework, to infer accurate and fine-grained 3D structures of objects in high-dimensional voxel space. Extensive experiments on large synthetic datasets and real-world Kinect datasets show that the proposed 3D-RecGAN++ significantly outperforms the state of the art in single view 3D object reconstruction, and is able to reconstruct unseen types of objects.

Recent advancements in deep learning opened new opportunities for learning a high-quality 3D model from a single 2D image given sufficient training on large-scale data sets. However, the significant imbalance between available amount of images and 3D models, and the limited availability of labeled 2D image data (i.e. manually annotated pairs between images and their corresponding 3D models), severely impacts the training of most supervised deep learning methods in practice. In this paper, driven by a novel design of adversarial networks, we have developed an unsupervised learning paradigm to reconstruct 3D models from a single 2D image, which is free of manually annotated pairwise input image and its associated 3D model. Particularly, the paradigm begins with training an adaption network via autoencoder with adversarial loss, which embeds unpaired 2D synthesized image domain with real world image domain to a shared latent vector space. Then, we jointly train a 3D deconvolutional network to transform the latent vector space to the 3D object space together with the embedding process. Our experiments verify our network's robust and superior performance in handling 3D volumetric object generation from a single 2D image.Existing works on 3D object reconstruction from 2D image(s) can be broadly categorized as two of the following: traditional methods without learning; deep learning based methods. 3D reconstruction without learning. The majority of traditional reconstruction methods based on SFM or SLAM [1,2] are subject to a dense number of views, and most of them rely on the hypothesis that features can be matched across views. 2D to 3D reconstruction models such as multi-view stereo [9, 10], space carving [11], multiple moving object and large scale structure from motion [3][4][5], have all demonstrated good performance in solving the 2D to 3D reconstruction problem. However these methods require high calibrated cameras and segmentation of objects from their background, which are less applicable in practice. Deep Neural Networks in 3D visual computing. Nowadays, by generating 3D volumetric data [12], prominent deep learning models such as the deep 2D convolutional neural networks can be naturally extended to learn 3D objects. Deep learning models have proven to have strong capabilities in learning latent representative vector space of 3D objects [12]. Multi-View CNN, Conv-DAE, Voxnet, Gift, T-L embedding, 3DGAN and so on, have uncovered great potential for solving retrieval, classification, 3D reconstruction problem, etc. on [13][14][15][16][17][18].In contrast to the vast amount of research and accomplishments in the field of 3D object classification and retrieval, there are fewer research and far less accomplished results on 3D object reconstruction. Recently, researchers began to utilize 3D deconvolutional neural network to generate 3D volumetric objects from 2D images, for instance, 3D- GAN[18] and T-L embedding [17] strive to learn a latent vector space representation of 2D images, and then transform it to gene...

Point cloud registration is to estimate a transformation to align point clouds collected in different perspectives. In learning-based point cloud registration, a robust descriptor is vital for high-accuracy registration. However, most methods are susceptible to noise and have poor generalization ability on unseen datasets. Motivated by this, we introduce SphereNet to learn a noise-robust and unseen-general descriptor for point cloud registration. In our method, first, the spheroid generator builds a geometric domain based on spherical voxelization to encode initial features. Then, the spherical interpolation of the sphere is introduced to realize robustness against noise. Finally, a new spherical convolutional neural network with spherical integrity padding completes the extraction of descriptors, which reduces the loss of features and fully captures the geometric features. To evaluate our methods, a new benchmark 3DMatchnoise with strong noise is introduced. Extensive experiments are carried out on both indoor and outdoor datasets. Under high-intensity noise, SphereNet increases the feature matching recall by more than 25 percentage points on 3DMatch-noise. In addition, it sets a new state-of-the-art performance for the 3DMatch and 3DLoMatch benchmarks with 93.5% and 75.6% registration recall and also has the best generalization ability on unseen datasets.

No abstract

We study the problem of recovering an underlying 3D shape from a set of images. Existing learning based approaches usually resort to recurrent neural nets, e.g., GRU, or intuitive pooling operations, e.g., max/mean pooling, to fuse multiple deep features encoded from input images. However, GRU based approaches are unable to consistently estimate 3D shapes given the same set of input images as the recurrent unit is permutation variant. It is also unlikely to refine the 3D shape given more images due to the long-term memory loss of GRU. The widely used pooling approaches are limited to capturing only the first order/moment information, ignoring other valuable features. In this paper, we present a new feed-forward neural module, named AttSets, together with a dedicated training algorithm, named JTSO, to attentionally aggregate an arbitrary sized deep feature set for multi-view 3D reconstruction. AttSets is permutation invariant, computationally efficient, flexible and robust to multiple input images. We thoroughly evaluate various properties of AttSets on large public datasets. Extensive experiments show AttSets together with JTSO algorithm 1 significantly outperforms existing aggregation approaches.

Odometry is of key importance for localization in the absence of a map. There is considerable work in the area of visual odometry (VO), and recent advances in deep learning have brought novel approaches to VO, which directly learn salient features from raw images. These learning-based approaches have led to more accurate and robust VO systems. However, they have not been well applied to point cloud data yet. In this work, we investigate how to exploit deep learning to estimate point cloud odometry (PCO), which may serve as a critical component in point cloud-based downstream tasks or learning-based systems. Specifically, we propose a novel end-to-end deep parallel neural network called DeepPCO, which can estimate the 6-DOF poses using consecutive point clouds. It consists of two parallel sub-networks to estimate 3-D translation and orientation respectively rather than a single neural network. We validate our approach on KITTI Visual Odometry/SLAM benchmark dataset with different baselines. Experiments demonstrate that the proposed approach achieves good performance in terms of pose accuracy.

We study the problem of efficient semantic segmentation for large-scale 3D point clouds. By relying on expensive sampling techniques or computationally heavy pre/postprocessing steps, most existing approaches are only able to be trained and operate over small-scale point clouds. In this paper, we introduce RandLA-Net, an efficient and lightweight neural architecture to directly infer per-point semantics for large-scale point clouds. The key to our approach is to use random point sampling instead of more complex point selection approaches. Although remarkably computation and memory efficient, random sampling can discard key features by chance. To overcome this, we introduce a novel local feature aggregation module to progressively increase the receptive field for each 3D point, thereby effectively preserving geometric details. Extensive experiments show that our RandLA-Net can process 1 million points in a single pass with up to 200× faster than existing approaches. Moreover, our RandLA-Net clearly surpasses state-of-the-art approaches for semantic segmentation on two large-scale benchmarks Semantic3D and Se-manticKITTI.