Search

Abstract From a 2D video of a person in action, human mesh recovery aims to infer the 3D human pose and shape frame by frame. Despite progress on video‐based human pose and shape estimation, it is still challenging to guarantee high accuracy and smoothness simultaneously. To tackle this problem, we propose a Video2mesh, a temporal convolutional transformer (TConvTransformer) based temporal network which is able to recover accurate and smooth human mesh from 2D video. The temporal convolution block achieves the sequence‐level smoothness by aggregating image features from adjacent frames. The subsequent multi‐attention transformer improves the accuracy due to its multi‐subspace for better middle‐frame feature representation. Meanwhile, we add a TConvTransformer discriminator which is trained together with our 3D human mesh temporal encoder. This TConvTransformer discriminator further improves the accuracy and smoothness by restricting the pose and shape in a more reliable space based on the AMASS dataset. We conduct extensive experiments on three standard benchmark datasets and show that our proposed Video2mesh outperforms other state‐of‐the‐art methods in both accuracy and smoothness.

Modeling fashion compatibility between different categories of items and forming personalized outfits have become important topics in recommender systems recently. However, item compatibility and outfit recommendation have been explored in perfect settings in the past, where high-quality images of items from the front view or user profiles are available. In this paper, we propose a new task called Complete The full-body Portrait (CTP) for real-world fashion images (e.g., street photos and selfies), which is able to recommend the most compatible item for a masked scene where the outfit is incomplete. Visual compatibility and personalization are the key points for accurate scene-based recommendations. In our approach, the former is accomplished by calculating the visual distance of the query scene and target item in latent space, while the latter is achieved by taking the body-shape information of the human subject into consideration. To obtain side information to train our model, ResNet-50, YOLOv3 and SMPLify-X models are adopted to extract visual features, detect item objects, and reconstruct a 3D body mesh, respectively. Our approach first predicts the missing item category from the masked scene, and then finds the most compatible items from the predicted category through computing visual distances at image level, region level and object level, together with measuring human body-shape compatibility. We conduct extensive experiments on two real-world datasets, Street2Shop and STL-Fashion. Both quantitative and qualitative results show that our model outperforms all baseline models.

Background"Horizontal strokes should be level and vertical strokes should be straight" is a common guideline in the teaching of Chinese handwriting. Measuring deviations in level horizontal and straight vertical strokes in students' Chinese handwriting is usually assessed manually. However, this task is time-consuming and may have inconsistent outcomes when judged by different people. In this paper, we aim to formulate a method to automatically evaluate the tilt and slant degrees of students' Chinese handwriting using digital handwriting tablets. Furthermore, we analyze the relationship between the tilt and slant features of students' Chinese handwriting and other demographic and handwriting features.MethodsFive hundred and ninety-one primary school students from grades 1 to 6 were recruited in Hong Kong. Before the assessment, a grid paper was attached to a digital handwriting tablet. The participants were then asked to copy 90 Chinese characters from a template to the grid paper. Their handwriting processes were recorded as two-dimensional points and then analyzed. The tilt and slant of the students' handwriting were calculated based on the inclination level of their horizontal and vertical strokes. Linear regressions between slant/tilt degree of the manuscripts and other handwriting features were performed. The students' demographic information was also explored.ResultsSlant was found to be significantly correlated to Gender (p < 0.001) and tilt×standard deviation of pen pressure (p < 0.001). Tilt was found to be significantly correlated to ground time (p < 0.001), slant (p < 0.001) and slant×special education need (p = 0.021).ConclusionsOur results demonstrate the relationship between slant, tilt and Chinese handwriting performance in primary school children. Slant and tilt can be adopted as an indicator in students' special education need diagnosis, as tilt level in the students' Chinese handwriting was related to ground time and slant× special education need, while slant is related to tilt×standard deviation of pen pressure and female students. These findings may also inspire ways to increase special education need students' writing speed.

Brain computer interfaces (BCIs) offer a broad class of neurologically impaired individuals an alternative means to interact with the environment. Many BCIs are "synchronous" systems, in which the system sets the timing of the interaction and tries to infer what control command the subject is issuing at each prompting. In contrast, in "asynchronous" BCIs subjects pace the interaction and the system must determine when the subject's control command occurs. In this paper we propose a new idea for BCI which draws upon the strengths of both approaches. The subjects are externally paced and the BCI is able to determine when control commands are issued by decoding the subject's intention for initiating control in dedicated time slots. A single task with randomly interleaved trials was designed to test whether it can be used as stimulus for inducing initiation and non-initiation states when the sensory and motor requirements for the two types of trials are very nearly identical. Further, the essential problem on the discrimination between initiation state and non-initiation state was studied. We tested the ability of EEG spectral power to distinguish between these two states. Among the four standard EEG frequency bands, beta band power recorded over parietal-occipital cortices provided the best performance, achieving an average accuracy of 86% for the correct classification of initiation and non-initiation states. Moreover, delta band power recorded over parietal and motor areas yielded a good performance and thus could also be used as an alternative feature to discriminate these two mental states. The results demonstrate the viability of our proposed idea for a BCI design based on conventional EEG features. Our proposal offers the potential to mitigate the signal detection challenges of fully asynchronous BCIs, while providing greater flexibility to the subject than traditional synchronous BCIs.