This paper focuses on the problems of point cloud deep neural networks in classification and segmentation tasks, including losing important information during down-sampling, ignoring relationships among points when extracting features, and network performance being susceptible to the sparsity of point cloud. To begin with, this paper proposes a farthest point sampling-important points sampling method for down-sampling, which can preserve important information of point clouds and maintain the geometry of input data. Then, the local feature relation aggregating method is proposed for feature extraction, improving the network's ability to learn contextual information and extract rich local region features. Based on these methods, the important points feature aggregating net (IPFA-Net) is designed for point cloud classification and segmentation tasks. Furthermore, this paper proposes the multi-scale multi-density feature connecting method to reduce the negative impact of point cloud data sparsity on network performance. Finally, the effectiveness of IPFA-Net is demonstrated through experiments on ModelNet40, ShapeNet part, and ScanNet v2 datasets. IPFA-Net is robust to reducing the number of point clouds, with only a 3.3% decrease in accuracy under a 16-fold reduction of point number. In the part segmentation experiments, our method achieves the best segmentation performance for five objects.

Hashtags are the keywords that describe the theme of social media content and have become very popular in influence marketing and trending topics. In recent years, hashtag prediction has become a hot topic in AI research to help users with automatic hashtag recommendations by capturing the theme of the post. Most of the previous work mainly focused only on textual information, but many microblog posts contain not only text but also the corresponding images. This work explores both image-text features of the microblog post. Inspired by the self-attention mechanism of the transformer in natural language processing, the visual-linguistics pre-train model with transfer learning also outperforms many downstream tasks that require image and text inputs. However, most of the existing models for multimodal hashtag recommendation are based on the traditional co-attention mechanism. This paper investigates the cross-modality transformer LXMERT for multimodal hashtag prediction for developing LXMERT4Hashtag, a cross-modality representation learning transformer model for hashtag prediction. It is a large-scale transformer model that consists of three encoders: a language encoder, an object encoder, and a cross-modality encoder. We evaluate the presented approach on dataset InstaNY100K. Experimental results show that our model is competitive and achieves impressive results, including precision of 50.5% vs 46.12%, recall of 44.02% vs 38.93%, and F1-score of 47.04% vs 42.22% compared to the existing state-of-the-art baseline model.

Gaussian process (GP) regression is widely used to find the extreme of a black-box function by iteratively approximating an objective function when new evaluation obtained. Such evaluation is usually made by optimising a given acquisition function. However, for non-parametric Bayesian optimisation, the extreme of the objective function is not a deterministic value, but a random variant with distribution. We call such distribution the maximum distribution which is generally non-analytical. To construct such maximum distribution, traditional GP regression method by optimising an acquisition function is computational cost as the GP model has a cubic computation of training data. Moreover, the introduction of acquisition function brings extra hyperparameters which made the optimisation more complicated. Recently, inspired by the idea of Sequential Monte Carlo method and its application in Bayesian optimisation, a Monte Carlo alike method is proposed to approximate the maximum distribution with weighted samples. Alternative to the method on GP model, we construct the maximum distribution within the framework of Student-t process (TP) which considers more uncertainties from the training data. Toy examples and real data experiment show the TP-based Monte Carlo maximum distribution has a competitive performance to the GP-based method.