The high intraclass variations and high interclass similarities make high spatial resolution (HSR) image interpretation very difficult. In this paper, we provide a feature selection method for scene classification, particularly to recognize the images from the industrial scene. The proposed method is based on scene overlapping rate and a quantitative feature assessment metric, considering the interclass distinguishing capabilities of features. Thus, the optimal feature representation for diverse scenes according to the discriminative characteristics of intraclass images can be selected. Specifically, an efficient scene recognition algorithm is presented to identify the easy-confused scenes from the dataset. An evaluation metric is also proposed metric to quantify the distinguishing capability of features with full statistical support. Experiments comparing 14 state-of-art classification methods on two challenging benchmark scene datasets (UCM and AID) show the effectiveness of the proposed method for HSR scene classification.
3D Point cloud, which is considered as the simplest and most efficient shape representation for 3D objects, has been widely used in various real-world applications such as virtual reality, autonomous driving and digital twin. Point cloud completion aims to predict the complete shape structure and recover faithful details given partial observation of an incomplete input. Unlike previous completion methods based on linear architectures, this paper presents a novel hierarchical architecture for point cloud completion which divides the completion process into several levels in a coarse-to-fine manner and significantly improves the network capacity for recovering local details. First, we exploit feature connections between encoded partial inputs and decoded recovery results at the same resolution by extracting multi-scale feature points, which can provide rich information for the following generation process. Second, in order to exploit the local geometric information and interpolate the extracted features points, we introduce cross-attention based generators into the decoding phase. The cross-attention based generator preserves produced structures from previous levels and incorporate the extracted feature points into each step of a progressive generation. Extensive experiments show that our method outperforms state-of-the-art completion approaches on popular PCN and ShapeNet55 datasets.
PointNet++ is a simple but effective network designed for point cloud processing. However, the accuracy of PointNet++ has been surpassed by many other methods, like DGCNN and Point Cloud Transformer. These methods are way heavier compared to PointNet++, which is not favorable for the deployment of real-world products. In this paper, we propose a module called HD projection layers that was inspired by nonlinear kernels used in support vector machines. The HD projection layers project the features of the point cloud into a higher dimension, increasing the linear separability and therefore relieving the burden on the classifier. Equipped with HD projection layers, we extended PointNet++ into a new network, HD-PointNet, which also involves many other improvements and better training techniques. Experiments show that the accuracy of HD-PointNet is competitive against other modern methods while using fewer computation resources.
The emergence of 3D point cloud analysis has brought about new opportunities and challenges in various fields such as autonomous driving, digital twins, and virtual reality. Accurate segmentation is crucial to 3D point cloud analysis, but challenges arise due to the lack of topological information, complex shapes, and sparsity and unevenness in point sampling. To address these problems, a novel point cloud segmentation network called PCSNet (Point Cloud Segmentation Network) has been proposed. PCSNet combines global and local features to determine the overall shape and detailed local information, respectively, through an encoder-decoder architecture that incorporates multi-scale feature fusion. The encoder progressively extracts local center points, fuses local features, and models global features with the transformer to construct multi-scale topological and semantic information. The decoder then recovers the original point cloud and incorporates multi-scale features by upsampling for accurate segmentation. PCSNet outperforms state-of-the-art point cloud segmentation approaches on two widely used benchmark datasets (ShapeNetPart and S3DIS).
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.