Point cloud is a widely used 3D data form, which can be produced by depth sensors, such as Light Detection and Ranging (LIDAR) and RGB-D cameras. Being unordered and irregular, many researchers focused on the feature engineering of the point cloud. Being able to learn complex hierarchical structures, deep learning has achieved great success with images from cameras. Recently, many researchers have adapted it into the applications of the point cloud. In this paper, the recent existing point cloud feature learning methods are classified as point-based and tree-based. The former directly takes the raw point cloud as the input for deep learning. The latter first employs a k-dimensional tree (Kd-tree) structure to represent the point cloud with a regular representation and then feeds these representations into deep learning models. Their advantages and disadvantages are analyzed. The applications related to point cloud feature learning, including 3D object classification, semantic segmentation, and 3D object detection, are introduced, and the datasets and evaluation metrics are also collected. Finally, the future research trend is predicted.

Dexterous in-hand manipulation poses significant challenges for a multi-fingered anthropomorphic hand due to the high-dimensional state and action spaces, as well as the intricate contact patterns between the fingers and objects. Although deep reinforcement learning has made moderate progress and demonstrated its strong potential for manipulation, it faces certain challenges, including large-scale data collection and high sample complexity. Particularly in scenes with slight changes, it necessitates the re-collection of vast amounts of data and numerous iterations of fine-tuning. Remarkably, humans can quickly transfer their learned manipulation skills to different scenarios with minimal supervision. Inspired by the flexible transfer learning capability of humans, we propose a novel framework called Progressive Transfer Learning (PTL) for dexterous in-hand manipulation. This framework efficiently utilizes the collected trajectories and the dynamics model trained on a source dataset. It adopts progressive neural networks for dynamics model transfer learning on samples selected using a new method based on dynamics properties, rewards, and trajectory scores. Experimental results on contact-rich anthropomorphic hand manipulation tasks demonstrate that our method can efficiently and effectively learn in-hand manipulation skills with just a few online attempts and adjustment learning in the new scene. Moreover, compared to learning from scratch, our method significantly reduces training time costs by 85%.

Tubular solder joint detection is an important and challenging issue in the industry, due to the illegible objects, rarely collected datasets and requiring high-precision and real-time performance for positioning and angle estimation. In this article, we propose an Attention integrated and Cross-spatial feature fused Rotation Network (ACR-Net) for tubular solder joint detection, which consists of a new feature extraction network named ECA-CSPDarknet44, a cross-spatial feature fusion network (CFFN), and a bin-based rotation detection network (BRDN). The proposed network can efficiently detect oriented tubular solder joints with high-precision and real-time performance. ECA-CSPDarknet44 with attention mechanism was presented, which can adaptively guide the network to learn important features of tubular solder joints, significantly improving the ability of feature extraction. By integrating multiscale global features and multiscale local region features, CFFN can enhance the network's ability to express the characteristics of tubular solder joints. Meanwhile, BRDN is proposed for oriented bounding box regression through a decoupling approach, which regresses target parameters efficiently and accurately with little increasing of model complexity. Finally, we establish a tubular solder joint dataset and conduct sufficient experiments to verify the effectiveness of our method. Our proposed ACR-Net achieves 98.8% mAP with 31.3 frames per second (FPSs) on the dataset, meeting the high-precision and real-time requirements of industrial systems.

Deep learning (DL) based object detection has achieved great progress. These methods typically assume that large amount of labeled training data is available, and training and test data are drawn from an identical distribution. However, the two assumptions are not always hold in practice. Deep domain adaptive object detection (DDAOD) has emerged as a new learning paradigm to address the above mentioned challenges. This paper aims to review the state-of-the-art progress on deep domain adaptive object detection approaches. Firstly, we introduce briefly the basic concepts of deep domain adaptation. Secondly, the deep domain adaptive detectors are classified into five categories and detailed descriptions of representative methods in each category are provided. Finally, insights for future research trend are presented.