Research Projects
Fashion elements mining and reasoning recommendation
Focused Research
Nowadays, clothing products occupy a vast e-commerce market share, and people's demand for clothing has also become more diversified. With the continuous development of computer vision and related technologies, new applications such as "smart fitting mirror" and smart clothing matching have kept emerging, bringing tremendous market potentials to the clothing industry. Thus it is necessary to develop a more accurate and clear identification of clothing attributes, which can contribute to the improvement and promotion of the above applications.
Researchers of this project have deeply explored the importance of landmark detection in clothing attribute recognition and found that using landmark joint learning can improve the performance of their model effectively. They adopted the data set from Fashion AI for research, and train their model to recognize the major eight attributes from the dataset, and finally increased the classification accuracy rate to over 90% on the validation data set (only includes pictures with real person model).
Knowledge graph refining and generation
Focused Research
At present, it is inefficient to manually extract venture capital information from the Internet and arrange them into categories such as the invested companies, investment institutions, amounts, fields, etc. Thus, this project aims to construct an algorithm that can automatically identify venture capital data from multiple sources, and find the most effective and authentic venture capital information so as to improve the accuracy and effectiveness of information extraction.
In order to extract venture capital financing events from news texts (usually natural language texts), the researchers mainly analyze data from multiple sources with different types of entities and extract venture capital concepts, such as entities, relationships, and attributes and arrange them in a structured form.
Key technology:
1) Recognize event triggers and event type based on deep learning technology;
2) Joint judgment technology for extracting Event Argument and judging its Argument Role;
Motion sensor for gesture classification
Focused Research
This project aims to assist users in using the dumbbell tutorial video on an app to learn dumbbell movements. The researchers construct an algorithm to score the user's dumbbell movements in the following three levels: perfect——the user completed an action correctly; good——the user completed the action, but with the wrong movement range or rhythm, keep up——the user's action is different from the tutorial action. It is required for the algorithm to be able to use the six-channel signals collected by the six-axis sensor to accurately grade the users' actions and ensure a low misjudgment rate. In addition, while taking the mobile phone's performance into account, the algorithm can quickly evaluate users' movements and produce results in real-time.
Using the acceleration data read by the sensor on the smart dumbbell, based on the method of statistical analysis, the researchers compare the relevant feature values of user's actions with these of the standard actions and using waveform fitting calculation to come out a comprehensive assessment of user's actions.
In the preliminary test, the average value of the algorithm's various indicators is pretty good with the accuracy rate of 80%, precision rate of 77.27%, recall rate 78.84%, and F1 Score rate 77.51%. The current algorithm has been launched in the new version of the app to test its performance more extensively.
Summary Scoring and Content Generation
Focused Research
Abstracts of academic papers can help relevant researchers quickly grasp the contextual knowledge in a specific field, while traditional abstract methods rely heavily on experts. The cooperative startup Sparrho resorts to curators and editorial teams to create short and simple summaries of difficult scientific papers and patents, costing a lot of workforces and material resources. Thus, how to effectively improve the efficiency and quality of abstracts is a scientific topic worth studying.
The researchers of this project propose a framework for the automatic abstraction of research papers, with labeled data generated manually by experts, through text extraction and generation, to make the automatic abstracts closer to manual abstracts. They mainly study the following aspects: 1) Extract key sentences (stage one) to assist experts and editors in quickly grasping the core content to save human resources. 2) Generate an automatic summary (stage two) to make the extracted sentence more accessible and coherent, fully automated.