Knowledge graph (KG) embedding aims to study the embedding representation to retain the inherent structure of KGs. Graph neural networks (GNNs), as an effective graph representation technique, have shown impressive performance in learning graph embedding. However, KGs have an intrinsic property of heterogeneity, which contains various types of entities and relations. How to address complex graph data and aggregate multiple types of semantic information simultaneously is a critical issue. In this article, a novel heterogeneous GNNs framework based on attention mechanism is proposed. Specifically, the neighbor features of an entity are first aggregated under each relation-path. Then the importance of different relation-paths is learned through the relation features. Finally, each relation-path-based features with the learned weight values are aggregated to generate the embedding representation. Thus, the proposed method not only aggregates entity features from different semantic aspects but also allocates appropriate weights to them. This method can capture various types of semantic information and selectively aggregate informative features. The experiment results on three real-world KGs demonstrate superior performance when compared with several state-of-the-art methods.

Head pose estimation (HPE) has wide industrial applications, such as online education, human–robot interaction, and automatic manufacturing. In this article, we address two key problems in HPE based on label learning and asymmetric relation cues: 1) how to bridge the gap between the better prediction performance of networks and incorrectly label pose images in the HPE datasets and 2) how to take full advantage of the adjacent poses information around the centered pose image. We reconstruct all the incorrect labels as a two-dimensional Lorentz distribution to tackle the first problem. Instead of directly adopting the angle values as hard labels, we assign part of the probability values ( soft labels) to adjacent labels for learning discriminative feature representations. To address the second problem, we reveal the asymmetric relation nature of HPE datasets. The yaw direction and pitch direction are assigned different weights by introducing the half at half-maximum of the Lorentz distribution. Compared with the traditional end-to-end frameworks, the proposed one can leverage the asymmetric relation cues for predicting the head pose angle in the incorrect label scenarios. Extensive experiments on two public datasets and our infrared dataset demonstrate that the proposed ARHPE network significantly outperforms other state-of-the-art approaches.

Fine-grained bird image classification (FBIC) is not only meaningful for endangered bird observation and protection but also a prevalent task for image classification in multimedia processing and computer vision. However, FBIC suffers from several challenges, such as bird molting, complex background, and arbitrary bird posture. To effectively tackle these challenges, we present a novel invariant cues-aware feature concentration Transformer (TransIFC), which learns invariant and core information in bird images. To this end, two novel modules are proposed to leverage the characteristics of bird images, namely, the hierarchy stage feature aggregation (HSFA) module and the feature in feature abstraction (FFA) module. The HSFA module aggregates the multiscale information of bird images by concatenating multilayer features. The FFA module extracts the invariant cues of birds through feature selection based on discrimination scores. Transformer is employed as the backbone to reveal the long-dependent semantic relationships in bird images. Moreover, abundant visualizations are provided to prove the interpretability of the HSFA and FFA modules in TransIFC. Comprehensive experiments demonstrate that TransIFC can achieve state-of-the-art performance on the CUB-200-2011 dataset (91.0%) and the NABirds dataset (90.9%). Finally, extended experiments have been conducted on the Stanford Cars dataset to suggest the potential of generalizing our method on other fine-grained visual classification tasks.

Human pose estimation (HPE) has many wide applications such as multimedia processing, behavior understanding and human-computer interaction. Most previous studies have encountered many constraints, such as restricted scenarios and RGB inputs. To mitigate constraints to estimating the human poses in general scenarios, we present an efficient human pose estimation model (i.e., EHPE) with joint direction cues and Gaussian coordinate encoding. Specifically, we propose an anisotropic Gaussian coordinate coding method to describe the skeleton direction cues among adjacent keypoints. To the best of our knowledge, this is the first time that the skeleton direction cues is introduced to the heatmap encoding in HPE task. Then, a multi-loss function is proposed to constrain the output to prevent the overfitting. The Kullback-Leibler divergence is introduced to measure the predication label and its ground truth one. The performance of EHPE is evaluated on two HPE datasets: MS COCO and MPII. Experimental results demonstrate that EHPE can obtain robust results, and it significantly outperforms existing state-of-the-art HPE methods. Lastly, we extend the experiments on infrared images captured by our research group. The experiments achieved the impressive results regardless of insufficient color and texture information.

Head pose estimation (HPE) is an indispensable upstream task in the fields of human-machine interaction, self-driving, and attention detection. However, practical head pose applications suffer from several challenges, such as severe occlusion, low illumination, and extreme orientations. To address these challenges, we identify three cues from head images, namely, critical minority relationships, neighborhood orientation relationships, and significant facial changes. On the basis of the three cues, two key insights on head poses are revealed: 1) intra-orientation relationship and 2) cross-orientation relationship. To leverage two key insights above, a novel relationship-driven method is proposed based on the Transformer architecture, in which facial and orientation relationships can be learned. Specifically, we design several orientation tokens to explicitly encode basic orientation regions. Besides, a novel token guide multi-loss function is accordingly designed to guide the orientation tokens as they learn the desired regional similarities and relationships. Experimental results on three challenging benchmark HPE datasets show that our proposed TokenHPE achieves state-of-the-art performance. Moreover, qualitative visualizations are provided to verify the effectiveness of the token-learning methodology.

2D Human pose estimation (HPE) has been widely used in the many fields such as behavioral understanding, identity authentication, and industrial automatic manufacturing. Most of the previous studies have encountered many constraints, such as restricted scenarios and strict inputs. To solve this problem, we present a simple yet effective HPE network called limb direction cues-aware network (LDCNet) with limb direction cues and differentiated Cauchy labels, which can efficiently suppress uncertainties and prevent deep networks from over-fitting uncertain keypoint positions. In particular, LDCNet suppresses the uncertainties from two aspects. (1) A differentiated Cauchy coordinate encoding method is designed to reveal the limb direction information among adjacent keypoints. (2) Jeffreys divergence is introduced as loss function to measure the prediction heatmap and ground-truth one. Positions of keypoints are perceived at the limb direction based deep network in an end-to-end manner. An extensive study on two benchmark data sets (i.e., MS COCO and MPII) illustrates the superiority of the proposed LDCNet model over state-of-the-art approaches.