Search

Neural rendering is an exciting topic injecting machine learning methodologies into the classical computer graphics rendering pipeline. Although recent works have achieved remarkable fidelity, discussion on how to enable it for interactive scenarios like video games seems to be lacking. Aside from an editable 3D model and UV-mapping, an interactive application will demand the neural rendering to handle animatable 3D content with interactive speed. This is currently a gap in neural rendering and our solution to this problem is a novel neural rendering pipeline involving a primitive named NeRFahedron. It localizes a NeRF field and as such effectively reduces the number of expensive network sampling operations to improve speed. Our pipeline involves tetrahedron rasterization, localized ray marching and near-surface particle sampling. The result is a method that can enable animatable content for neural rendering with interactive speed, which has been shown to be competitive in rendering animation. We will also showcase its ability to enable interactive applications via a real-time demo.

The increasing number of suicide events in recent years has set off an alarm in society. To prevent suicides, it is important to develop platforms with associated intelligent algorithms, such as those for early suicide ideation detection (SID). In general, entering people’s lives to obtain indicative information for effective SID is prohibitive due to the possible risk of privacy invasion. Social media posts provide valuable information about users’ activities, indicating important hints toward SID in a non-intrusive manner. Although multiple datasets have been collected from social media platforms for efficient SID, they either neglect many suicidal posts by searching for pre-defined keywords, or contain limited information about suicide ideation. In this paper, a newly collected dataset, which expands the coverage of suicidal posts with more fine-grained annotations of suicide risk levels and suicide triggers compared with existing datasets, is presented. Benchmarking results by a popular deep-learning model are analysed to validate the reliability and potential of the collected dataset. Furthermore, multiple application scenarios of the dataset are discussed. The proposed dataset is expected to enhance the research on suicide ideation and behaviours, and have a strong impact on SID and more importantly, suicide prevention to save invaluable lives.

Have you ever picked up a stuffed toy and pretended to play with it in your childhood? We are motivated by the novel use of stuffed toys in enhancing extended reality interaction. A key goal of extended reality is to induce the feeling of presence in its users. Naturally mapped control interface has been shown to enhance presence. The literature also indicates that a high degree of freedom tracking is important to extended reality. Based on these observations, we show that a free-form naturally mapped control interface is well-motivated via a theoretical contextualization. We explore the possibility of building such a controller in the form of stuffed toys. To realize stuffed toys as controllers, a novel soft-pose estimator empowered by cage-based deformation is proposed. It is shown to be effective in tracking the poses and deformations of real soft objects even by training with synthetic data only. Three gameplay prototypes are developed to demonstrate that interactive play can be enabled by the soft-pose estimator. They also form the basis for two user studies that validate the success of tracking stuffed toys with the soft-pose estimator for interactive play.

As a high-impact educational practice, service-learning has demonstrated success in positively influencing students’ overall development, and much work has been done on investigating student learning outcomes from service-learning. A particular direction is to model students’ learning outcomes in the context of their learning experience, i.e., the various student, course, and pedagogical elements. It contributes to a better understanding of the learning process, a more accurate prediction of students’ attainments on the learning outcomes, and improvements in the design of learning activities to maximize student learning. However, most of the existing work in this area relies on statistical analysis that makes assumptions about attribute independence or simple linear dependence, which may not accurately reflect real-life scenarios. In contrast, the study described in this paper adopted a neural network-based approach to investigate the impact of students’ learning experience on different service-learning outcomes. A neural network with attention mechanisms was constructed to predict students’ service-learning outcomes by modeling the contextual information from their various learning experiences. In-depth evaluation experiments on a large-scale dataset collected from more than 10,000 students showed that this proposed model achieved better accuracy on predicting service-learning outcomes. More importantly, it could capture the interdependence between different aspects of student learning experience and the learning outcomes. We believe that this framework can be extended to student modeling for other types of learning activities.