Although 3D point cloud classification neural network models have been widely used, the in-depth interpretation of the activation of the neurons and layers is still a challenge. We propose a novel approach, named Relevance Flow, to interpret the hidden semantics of 3D point cloud classification neural networks. It delivers the class Relevance to the activated neurons in the intermediate layers in a back-propagation manner, and associates the activation of neurons with the input points to visualize the hidden semantics of each layer. Specially, we reveal that the 3D point cloud classification neural network has learned the plane-level and part-level hidden semantics in the intermediate layers, and utilize the normal and IoU to evaluate the consistency of both levels' hidden semantics. Besides, by using the hidden semantics, we generate the adversarial attack samples to attack 3D point cloud classifiers. Experiments show that our proposed method reveals the hidden semantics of the 3D point cloud classification neural network on ModelNet40 and ShapeNet, which can be used for the unsupervised point cloud part segmentation without labels and attacking the 3D point cloud classifiers.

Event cameras have emerged as a promising vision sensor in recent years due to their unparalleled temporal resolution and dynamic range. While registration of 2D RGB images to 3D point clouds is a long-standing problem in computer vision, no prior work studies 2D-3D registration for event cameras. To this end, we propose E2PNet, the first learning-based method for event-to-point cloud registration. The core of E2PNet is a novel feature representation network called Event-Points-to-Tensor (EP2T), which encodes event data into a 2D grid-shaped feature tensor. This grid-shaped feature enables matured RGB-based frameworks to be easily used for event-to-point cloud registration, without changing hyper-parameters and the training procedure. EP2T treats the event input as spatio-temporal point clouds. Unlike standard 3D learning architectures that treat all dimensions of point clouds equally, the novel sampling and information aggregation modules in EP2T are designed to handle the inhomogeneity of the spatial and temporal dimensions. Experiments on the MVSEC and VECtor datasets demonstrate the superiority of E2PNet over hand-crafted and other learning-based methods. Compared to RGB-based registration, E2PNet is more robust to extreme illumination or fast motion due to the use of event data. Beyond 2D-3D registration, we also show the potential of EP2T for other vision tasks such as flow estimation, event-to-image reconstruction and object recognition. The source code can be found at: https://github.com/Xmu-qcj/E2PNet.

Abstract Event camera is a new vision sensor that produces independent asynchronous responses to each pixel's change of illumination intensity. The unique principle of event camera has many advantages over traditional cameras, such as low latency, high temporal resolution, and high dynamic range (HDR). These advantages make event camera ideal for dealing with high speed, HDR visual tasks, especially in automatic driving scenes. In this study, we propose an image generation network named Event‐based attention generative adversarial networks (EAGAN), which simultaneously deals with optical flow and depth estimation of monocular event camera data. In addition to the innovative network architecture and loss function suitable for depth estimation, we are also the first to process incomplete training data to obtain more dense and uniform prediction results. Experiments on the multi‐vehicle stereo event camera dataset show that our EAGAN is competitive on the depth estimation task and achieves the state‐of‐the‐art effect in the optical flow estimation task.

Similarity measure of cross-domain descriptors (2D descriptors and 3D descriptors) between 2D image patches and 3D point cloud volumes provides stable retrieval performance and establishes the spatial relationship between 2D and 3D space, which plays the potential applications in geospatial space, such as 2D and 3D interaction of remote sensing, Augmented Reality (AR) and robot navigation. However, the mature handcrafted descriptors of 2D image patches and 3D point cloud volumes are extremely different, resulting in the huge challenge for 2D image patch and 3D point cloud volume matching. In this paper, we propose a novel network which combines both unified descriptor training and descriptor comparison function training for 2D image patch and 3D point cloud volume matching. First, two feature extraction networks are applied for jointly learning the local descriptors for 2D image patches and 3D point cloud volumes, respectively. Second, a fully connected network is introduced to compute the similarity between 2D descriptors and 3D descriptors. Motivated by the successful indicator system on evaluating 2D patch feature representation, we use the false positive rate at 95% recall (FPR95) and precision based on cross-domain descriptors as the measured metric. The experimental results show that our proposed network achieve state-of-the-art performance in the matching of 2D image patches and 3D point cloud volumes.

Different domain image sensors or imaging mechanisms provide cross-domain images when sensing the same scene. There is a domain shift between cross-domain images so that the image gap between different domains is the major challenge for measuring the similarity of the feature descriptors extracted from different domain images. Specifically, matching ground camera images and unmanned aerial vehicle (UAV) 3-D model-rendered images, which are two kinds of extremely challenging cross-domain images, is a way to establish indirectly the spatial relationship between 2-D and 3-D spaces. This provides a solution for the virtual-real registration of augmented reality (AR) in outdoor environments. However, during matching, handcrafted descriptors and existing learning-based feature descriptors limit the rendered images. In this letter, first, to learn robust and invariant 128-D local feature descriptors for ground camera and rendered images, we present a novel network structure, SiamAM-Net, which embeds the autoencoders with an attention mechanism into the Siamese network. Then, to narrow the gap between the cross-domain images during the optimizing of SiamAM-Net, we design an adaptive margin for the loss function. Finally, we match the ground camera-rendered images by using the learned local feature descriptors and explore the outdoor AR virtual-real registration. Experiments show that the local feature descriptors, learned by SiamAM-Net, are robust and achieve state-of-the-art retrieval performance on the cross-domain image data set of ground camera and rendered images. In addition, several outdoor AR applications also demonstrate the usefulness of the proposed outdoor AR virtual-real registration.

This paper presents a novel virtual-real registration approach for augmented reality (AR) in large-scale outdoor environments. Essentially, it is a pose estimation for the mobile camera images (ground camera images) in 3D model recovered by Unmanned Aerial Vehicle (UAV) image sequence via Structure-From-Motion (SFM) technology. The approach considers to indirectly establish the spatial relationship between 2D and 3D space by inferring the transformation relationship between the ground camera images and the UAV 3D model rendered images. Specifically, the proposed approach can overcome the positioning errors, which are deterioration and drift in the GPS, and deviation of orientation. The experimental results demonstrate the possibility of the proposed virtual-real registration approach, and show that the approach is robust, efficient and intuitive for AR in large-scale outdoor environments.