The generative pre-training transformer (GPT) has been demonstrated to be remarkably effective in natural language processing. Motivated by advances in GPT, a geometric matching network with a generative pre-training transformer (GMNet) is proposed in this paper to provide high-quality automatic 3D object tracking for intelligent systems. Specifically, the characteristics of GMNet can be summarized as follows. First, a geometry-rebuild pre-training strategy is proposed to enhance the perception of incomplete geometry. Under this training strategy, the network is pre-trained on point cloud completion datasets to learn shape-inference capabilities and obtain a robust latent representation for tracking. After pre-training, the network is fine-tuned with the tracking data to achieve high-precision object tracking. Second, a shape reasoning module (SRM) with generative pre-training is designed to perform point cloud completion, aiming to refine geometric features by leveraging shape perception and reasoning abilities learned from geometry-rebuild pre-training. Third, an adaptive gated feature matching module (FMM) is designed to perform adaptive feature matching via a gate mechanism, which predicts the object’s location based on adaptively matched information derived from cosine similarity and SRM. Furthermore, extensive experiments demonstrate that the proposed GMNet outperforms state-of-the-art methods on KITTI, nuScenes, and Waymo open benchmarks.

Nonnegative matrix factorization (NMF) is a powerful tool for signal processing and machine learning. Geometrically, it can be interpreted as the problem of finding a conic hull, which contains a cloud of data points and is embedded in the positive orthant. The separability assumption posits that the conic hull can be spanned by a small subset of the columns of the input data matrix. This assumption is equivalent to the 1-sparse condition. Many extreme-rays-based NMF methods are essentially based on the 1-sparse condition. However, the separability assumption or 1-sparse condition may not always be guaranteed for real applications. By analyzing the mathematical connection between the extreme-rays representation and the half-hyperplanes representation of a conic hull, we propose three novel NMF algorithms (i.e., HICHF, EnhancedHICHF, and ExtendedHICHF) based on the half-hyperplane identification. These algorithms can be efficiently implemented via eigenvalue decomposition (EVD). In contrast to the conventional extreme-rays-based NMF methods, the proposed methods can achieve better performance for the nonseparable NMF problems, where the 1-sparse condition is not well satisfied. Furthermore, the proposed algorithms are simple, yet efficient and more robust. Experiments on both synthetic data and real-world parts-based learning data, such as hyperspectral unmixing and facial parts learning, verify that the proposed algorithms considerably outperform the state-of-the-art algorithms.

Point cloud completion concerns the inference of the completed geometries for real-scanned point clouds that are sparse and incomplete due to occlusion, noise, and viewpoint. Previous methods usually learn a one-shot partial-to-complete mapping, which is incapable of generating fine structure details for the complex point cloud distributions. In this paper, a Hierarchical Geometry Generation point completion Network (HGG-Net) is proposed to hierarchically generate the fine-grained completed point cloud with a skeleton-to-details strategy, which consists of three fundamental modules, namely Transformer-enhanced Feature Encoder (TFE), Multi-level Geometry Representation Decoder (MGRD), and Hierarchical Dynamic Geometry Generator (HDG). Specifically, TFE first extracts geometry features of the incomplete input and obtains a coarse prediction via self-attention mechanism and edge convolution. Second, MGRD obtains the multi-level decoded geometry representations by Geometric Interactive Transformer (GIT) and Channel-Attention-based Geometry Features Fusion (CAGF), where GIT is proposed to decode the complete prompt by capturing the semantic relationship between geometry features of the incomplete and the decoded complete objects, and CAGF aims to fuse them for the high-quality representation. Third, HDG generates the complete points hierarchically from skeleton to details based on the Dynamic Graph Attention mechanism. Qualitative and quantitative experiments demonstrate that the proposed HGG-Net outperforms state-of-the-art methods on several point cloud completion datasets. Our code is available at https://github.com/haalexx/HGGNet.

Automatic segmentation of 3D dental models into individual teeth is an important step in orthodontic computer-aided design (CAD) systems. However, most existing methods rely on single-view dental models and ignore the intrinsic relationships between upper and lower dental models, hindering the handling of complex tooth structures. In this paper, a collaborative learning framework with coupling graph Transformers (CGT-CLF) is proposed for automatic tooth segmentation on 3D dental models. The framework collaboratively learns geometric features of both upper and lower dental models, capturing their interactivity and complementarity by facilitating interaction between graph-Transformer encoders to improve segmentation of complex and diverse teeth. Specifically, CGT-CLF consists of three key components as follows: First, a graph embedding-based boundary perception module (GEBPM) is developed to aggregate fine-grained geometric features within the neighborhood graph domain, enhancing the network's ability to perceive and distinguish intricate tooth boundaries. Then, coupling geometric Transformers are designed to capture the intrinsic relationships of pair-wise dental models by promoting the exchange of relevant information to gain a comprehensive understanding of the overall tooth structure, allowing for better identification of adjacent teeth with similar appearances. Finally, a collaborative cross-scale feature fusion (CCFF) strategy is utilized to obtain interactive and complementary information by modeling the inter-relationships between dual-stream features. Experimental results on a clinical dental model dataset demonstrate that the proposed CGT-CLF framework outperforms state-of-the-art methods, delivering superior segmentation performance.

Accurate object detection in remote sensing images (RSIs) is of great significance for various applications such as environmental monitoring and agricultural production. However, it is a challenging task mainly due to the complex backgrounds and scale diversity of geospatial objects. In this letter, a Cross-Scale Hybrid Gaussian Attention Network (CSHGANet) is proposed for accurate object detection in RSIs, and it consists of two main components as follows. First, hybrid Gaussian attention is designed to learn the interrelationships between channels and spatial locations of features, which can focus on geospatial objects and reduce the interference of complex backgrounds in RSIs. Then, a cross-scale feature aggregation module is developed to adaptively fuse multi-scale attention feature maps to capture more rich and discriminative feature representations, so as to better handle scale variations of remote sensing objects. Extensive experiments on two public datasets (i.e., NWPU VHR-10 and RSOD) show that the proposed CSHGANet outperforms state-of-the-art object detection methods, achieving mean average precision (mAP) scores of 95.53% and 98.61%, respectively.

Accurate segmentation of 3D dental models derived from intra-oral scanners (IOS) is one of the key steps in many digital dental applications such as orthodontics and implants. However, it is difficult to accurately segment individual teeth and gums in 3D dental models due to the following problems: 1) the shape and appearance of adjacent teeth are very similar, which is easy to be misidentified; 2) the boundary between teeth and gums is often indistinct, especially in orthodontic patients with abnormalities such as missing and crowded teeth. To solve such problems, a Dual-Branch Geometric Attention Network (DBGANet) for 3D tooth segmentation is proposed, which can capture tooth geometric structure and detailed boundary information from multi-view geometric features encoded by 3D coordinates and normal vectors. The framework contains two branches, i.e., C-branch and N-branch. First, centroid-guided separable attention is designed in the C-branch to learn global context information by modeling the spatial dependencies of tooth point clouds, which can capture the overall geometric structure of teeth to better distinguish adjacent teeth with similar appearance. Then, Gaussian neighbor attention is designed in the N-branch to encode normal vectors to highlight detailed differences between geometric features at different points, which helps to refine the boundaries of teeth and gingiva for more accurate and smooth tooth segmentation. Extensive experiments on the real-patient datasets of 3D dental models demonstrate that the proposed DBGANet significantly outperforms state-of-the-art methods.

Accurate and rapid diagnosis of COVID-19 using chest X-ray (CXR) plays an important role in large-scale screening and epidemic prevention. Unfortunately, identifying COVID-19 from the CXR images is challenging as its radiographic features have a variety of complex appearances, such as widespread ground-glass opacities and diffuse reticular-nodular opacities. To solve this problem, we propose an adaptive attention network (AANet), which can adaptively extract the characteristic radiographic findings of COVID-19 from the infected regions with various scales and appearances. It contains two main components: an adaptive deformable ResNet and an attention-based encoder. First, the adaptive deformable ResNet, which adaptively adjusts the receptive fields to learn feature representations according to the shape and scale of infected regions, is designed to handle the diversity of COVID-19 radiographic features. Then, the attention-based encoder is developed to model nonlocal interactions by self-attention mechanism, which learns rich context information to detect the lesion regions with complex shapes. Extensive experiments on several public datasets show that the proposed AANet outperforms state-of-the-art methods.