Researched in 3D vision, 3D generation, and imitation learning under the supervision of Prof. Xiaolong Wang.

· Researched in 3D hand pose estimation from depth images and binocular images· Made use of Kinect, an interactive equipment, to gather finger depth map, and Leap Motion to collect finger binocular images, and estimated three-dimensional coordinates of fingertips by using stereopsis· present a light and efficient approach named Latent2DFusion network for fast andaccurate hand pose estimation from a single depth image.· Achieved state-of-the-art accuracy on three public hand pose datasets and outperformed all top-ranked approaches by a large margin in terms of inference speed (nearly thousand frames per second) and model size (less than 10 MB).

· Researched image quality assessment, mainly responsible for blind face image quality assessment· Evaluated face image quality by using statistical approaches based on standard human face identification photos and deep learning approaches based on the idea of learning to rank and Siamese Convolutional Neural Networks; Combined the two methods flexibly, regarded deep neural networks as strong feature extractors, and conducted One-class Classification in the deep feature space of input human faces, successfully solving the problem of automatic face image quality predictions· Successfully applied the model for human face quality assessment into product line

· Self-studied Hadoop and Web crawler to collect three million pieces of flower images; Made fine-grained classification with them through deep learning and established a high efficient and accurate processing pipeline; Made successful market appearance of the classifier· Revised Softmax loss, added negative samples that correspond to some uncertain flower images, aiming at better compressing the false recognized sample space while reducing the system error rate. Adjusted gradient descent optimization algorithm to perfectly fit the new loss function, thus enabled the Convolutional Neural Network to make use of negative samples to accomplish better study · Raised the accuracy of top-3 from 59% to 61%, and that of top-5 from 75% to 78%, and ensured the recall rate at 90%