Baiqiang Xia is an experienced AI Scientist specialized in computer vision and deep learning, particularly in 2D/3D face recognition, hand and body motion recognition, visual object recognition, and eye-tracking. Baiqiang holds a PhD in 3D Face Recognition and has a strong background in academia. He has been awarded several recognitions for both his research work and industry applications.
As an AI scientist, Baiqiang has applied his knowledge on various academic and industrial projects, including healthcare, media, and heavy industry, to mention a few. Typically the projects he has embarked on have required visual recognition of different characteristics, such as identifying a person’s gender, age, expression, pose, gaze, motion, or locating the objects in the visual scene.
Industrial projects have a specific context
Baiqiang’s work requires a lot of testing and evaluating different approaches for visual recognition. There are usually multiple AI solutions in the research literature for a specific problem. These AI models have well-documented speed and accuracy reports based on benchmark datasets. This type of resources are valuable to the project work, in a sense of providing the scope of promising computer vision models.
– However, typically a model that performs best on a benchmark (mock up) dataset usually is not the best choice for a specific project. Finding the right model requires a lot of customization and testing, Baiqiang explains.
For example, face detection usually makes the entry point for any facial analysis. The domain literature provides many face detectors, from the classical hand-crafted features based models (e.g. Viola-Jones, HoG Face Detector), to the very recent deep learning models (e.g MMOD, MTCNN).
The MTCNN face detector makes the state-of-the-art on many benchmark datasets, while the 20-year-old Viola-Jones face detector might be a better choice if the faces in the image data are close to camera, near-frontal and with sufficient illumination. Despite its age, Viola-Jones face detector works much faster and relatively accurately under this scenario.
Curiosity brought to computer vision
Baiqiang says he got interested in computer vision out of curiosity. We humans perform visual recognition naturally and effectively in our daily lives. The questions about the underlying mechanism have always been intriguing, and largely unanswered – how are these functionalities performed and can we apply a similar methodology with computers to achieve similar functionality?
Computer vision provides a unique and powerful way towards these goals, but many things still require further research.
– Think about wrinkles on your face. We know that we can tell effectively young people from old people, and wrinkles are strong cues for distinguishing the two. But how much are wrinkles correlated to age? Can we incorporate this idea with computer vision and machine learning techniques, to see how much accuracy could the wrinkles provide? For me these questions are always fascinating to think of and work on, Baiqiang says.
Anything about data is important
Data management is always at the center of AI projects. In AI projects, the workload concerning data management is usually under-estimated, which might add challenges and risks to the success of the project.
On one hand, we always need to learn from data to build AI models. Data collection, preprocessing and labelling can take a considerable amount of time, from one third to more than two thirds of the project time. The lack of clean and labelled data, or delay of data preparation, are noticeably common issues in computer vision projects.
On the other hand, there are rising public concerns about data privacy and intellectual property – which is good for the sustainable development of the AI-ecosystem.
– It is crucial to make the project work comply with the latest data privacy guidelines, and satisfy the customers data privacy requirements, to minimize the legal risks and build the long-term trust with clients. We need to always prioritize questions on how data should be collected, stored, transferred, analyzed, reported and destroyed, especially in sensitive domains like face analysis and healthcare, Baigiang comments.
Computer vision technologies are advancing fast
When Baiqiang talks about his work, he highlights the rapid speed in which AI and deep learning technologies advance, particularly in visual objects and face recognition domains. Keeping up to date with the latest technologies means you need to continuously track the latest research achievements. Baiqiang follows closely scientific conferences such as the computer vision conference CVPR and the International Conference on Machine Learning ICML, and recent updates in Arxiv.
Keeping a track isn’t always easy according to Baiqiang:
– As AI publications are exploding in number, I often need to prioritize which papers to read. I try to select milestone papers, which will be used as a basis of further research and that have been tested with many people. A good example like this is the ResNet paper, that came already back in 2015.
At the same time, Baiqiang follows topics he is personally more interested in, for instance few shot learning, which means creating AI with limited data and edge computing, where AI solutions are deployed on devices with limited computational resources. He believes AI will have a second boom when it lowers down the dependency on data, and reaches pervasively to the edge devices.
Human communication matters
In client work, Baiqiang wants to always emphasize communication: humans work with humans and they form the bridge in any technology development projects. Within our clients, Baiqiang is praised for his positive thinking and being always open to communication.
– I seek to find the clients’ needs and concerns and get them fixed on time. The non-technical part is of high importance for the project to succeed. You need trust between people to develop something in collaboration. As long as you keep the communication channel open, you can form a sustainable, long-term collaboration, Baiqiang explains.
Chinese chess and football
When not at work Baiqiang likes to play Chinese Chess – which is not the same Chess typically known in Finland. As a strategy game, this one too requires concentration and persistence. For sports, Baiqiang keeps fit with football every week.
Baiqiang is also an excellent cook: Silo.AI team got to experience Baiqiang’s amazing dumplings at our Christmas Party!
Favorite Silo.AI value?
– My favorite value is ‘Keep Learning’. I appreciate growth, both from the personal side, and from the company side. Learning is very important in my field and I try to participate in Silo.AI’s Learning Lab activities such as All-hands and Research Club and contribute by sharing my knowledge too as often as I can, Baiqiang concludes.
Want to join Silo.AI as Baiqiang’s colleague?
We are especially looking for Computer Vision experts to our offices in Helsinki and Turku to solve real-life cases using latest CV & ML techniques. .