MRI technology can obtain the brain avatar of patients from multiple modes and angles, which greatly improves the diagnostic efficiency of brain tumor patients. However, the MRI data of brain tumor patients is usually not enough to train the deep learning based segmentation model for sufficient performance. This paper proposes a simple and effective method to augment the data to improve the image segmentation effect of UNet++ model on brain glioma. To enhance the illumination robust of MRI image, we bring in the max filter and min filter for image augmentation, and evaluate the segmentation effect of the UNet++ model using the original dataset and the augmented dataset on glioma, and evaluate the experimental results. We used the public datasets MICCAI_ BraTS_20 to evaluate, The WT PPV values of the two models were 0.9233 and 0.9427 respectively, and other evaluation indicators were also improved to varying degrees. In summary, it is concluded that our proposed simple augmentation method could effective augment the brain glioma MRI segmentation performance of UNet++ model on glioma, especially in the limited data.
The computer vision field is quickly developing, including multiple object tracking, as the big data age approaches. The majority of the effort is focused on tracking methods while less attention is paid to the most important aspect, data. After an analysis of existing datasets, we find that they commonly ignore the breakpoint problem in tracking and have low image quality. Thus we present the dataset named Multiple Object Tracking on Horizontal view (MOT-H). MOT-H is meticulously annotated on crowded scenes from the horizontal view, with the primary goal of proving anti-jamming performance against complicated occlusion or even complete occlusion. The breakpoint issue is emphasized, which means the target object temporarily leaves the scene and returns after a while. The proposed MOT-H dataset has 10 sequences, 20,311 frames, and 337,440 annotation boxes in total, with all pictures having the resolution of 3840 × 2160 and being filmed at 30 frames per second (fps). We establish a fair benchmark for the future object tracking method development. The whole dataset can be found at: https://drive.google.com/drive/folders/1SCUJAdbqXQStyV-F2M9UyGfsuCaxR73a?usp=sharing.
Part segmentation is one of the important tasks in 3D shape analysis. Prior works mostly rely on complex local modeling to learn features directly from raw mesh or point cloud. We proposed a simple, fast and robust approach for 3D shape part segmentation without sophisticated local geometric modeling and ingenious networks. Our main idea is to learn more discriminative features from different geometric descriptors by sparse convolution and attention mechanism. Specifically, an effective shape representation block, consists of a features embedding module, two attention module. The coordinates are fed into the features embedding module to produce embedding vectors. The attention head to learn the intrinsic relations between different descriptors and produce more informative feature maps. In the lower branch, We use some geometric descriptors as local features. And a classification head predicts the corresponding label. Conditional Random Field (CRF) is applied to optimize the segmentation of the network. The experiment results on Princeton Shape Benchmark(PSB) demonstrate that our architecture outperforms different methods, with higher accuracy, lower complexity and faster speed.
Accurate traffic prediction has a wide spectrum of benefits for the society. As Electric Vehicles (EV) are being used more frequently, an accurate EV Charging Station availability forecast becomes necessary. These forecasting models can help alleviate congestion at charging stations and maneuver EV drivers to an ideal location. There have been many data-driven models applied to similar problems related to traffic prediction [1]. Despite numerous efforts to ameliorate already established models and to innovate new models, the complex topological structure of the road networks and the variations of temporal data hinder the models from achieving a good prediction. As for the most classic Recurrent Neural Network (RNN) only extracts time-series information, its lack of information related to spatial dependencies cause a significant loss in long-term accuracy. The Graph Neural Network (GNN) can perform well in extracting spatial dependencies but cannot process time-series information well. Combining the Graph Convolutional Network (a variant of GNN) and the Transformer, we seek to efficiently capture both spatial and temporal data related EV Charging Station Availability. The model is tested for its performances using the Dundee City dataset. And the result reflects that our model surpassed an accuracy of 80% and attained more accurate predictions than the classic baselines.
Proteins are critical for lives. They not only build 10%-35% of our body tissues, but also can be used to understand the structures of different viruses, and then help us to explore effective vaccines. Hence, predicting new protein structures is very important for human health. However, the structure of protein is complicated. Exploration using human experiments is cost-consuming. Recently, artificial intelligence (AI) technology, such as imitation learning and reinforcement learning (RL), has been rapidly developed and significantly improved the efficiency in many different domains. In this project, we will try to use RL to solve the protein folding structure prediction problem.
