Liang and his team are looking to enhance how computer vision models are reliably deployed in the real world, exploring the current limitations of computer vision technologies, particularly in driverless cars that are typically trained to operate in perfect conditions. However, reality is far from perfect.
Liang hopes to explore how computers “see” and adapt in complex and challenging environments, such as navigating through dense smoke during bushfires or partially flooded roads.
Liang is excited to leverage the opportunities this ongoing fellowship and funding provides to form global partnerships, collaborate with computer vision experts around the world and contribute long-term to greater vehicle safety for drivers and passengers.
Computers vs Humans – who’s the better driver?
Liang and his team are teaching computers to recognise increasingly complex images and videos, much like humans do. For instance, when a computer encounters an image of a pedestrian, it must assess the pedestrian’s location and distance from the vehicle. When a human drives a car, they can use their complex decision-making abilities to assess multiple factors and make multiple predictions and decisions simultaneously. Using computer vision the goal is to combine images and videos to train the computer to assess multiple factors at once to safely guide the car. Beyond identifying pedestrians, cameras on the car will inform how the computer recognises traffic light signals, road lines, cars and animals that may come onto a road. But what if we add more information?
Not every puddle on a road looks the same
Liang’s new research will focus on enhancing the reliability of autonomous vehicles in challenging environments. The most accurate computer vision systems are trained using data captured in controlled systems, where lighting, viewpoints, and resolution are ideal. Liang sees an opportunity to analyse training data, test environments and model uncertainty levels to reflect the real world and its unpredictability.
“Just like a human you can learn in the classroom but sometimes it’s important to go out and apply your knowledge in the real world, and that’s what we’re trying to do when training computers to drive cars.”
– Dr Liang Zheng
There are many challenges to using computer vision in unpredictable environments and so far, the most accurate systems have been to control the environment to improve accuracy. For example, at Sydney Airport, computer vision is being used in face recognition systems. Being a well-lit, controlled environment, can provide prompts to people to take off their glasses or stand in the right spot, making it easy for the video and images to be processed and analysed. But in real-world scenarios, we cannot always control the environment.
Liang’s work aims to push computer vision beyond these controlled environments, training it to understand the test environments and select the best system configuration even in the face of uncertainty.
Chasing the last 1%
With computer vision, it’s all about accuracy, down to the final one per cent. The real challenge lies in chasing down that last one per cent – the difference between a system that works in theory and one that is safe and effective to use in the real world.
“In autonomous vehicles, accuracy of perception is critical to keep us safe. There’s no space for inaccuracy on the road.”
Liang’s research is also exploring AI systems that not only give you an answer but also an indication of the level of confidence.
What’s next for AI?
Liang hopes to see the impact of this research and algorithms in real-world products and systems of industry partners like Seeing Machines who he has been collaborating with through an ARC Linkage Project. When asked about the role of AI in our daily lives, he says,
“In my opinion, I don’t think AI would evolve to the level that it will replace humans. So, I’m not that worried, it’s not like the sci-fi movies. But what I’m worried about is trying to let those AI systems benefit our everyday lives. Making our lives more efficient and safer.”
Impact of the ARC Future Fellowship
Liang is grateful for the funding of the fellowship, for it will allow him to focus on his research and building a team. Over the next four years of funding, it provides opportunities to travel and collaborate with people and organisations from around the world in the US, UK, across Australia and within ANU. It will also support three PhD students and a Postdoctoral role.
“One of the things I’m most excited about is that this grant gives me a lot of opportunities to connect with people and collaborate across the field of computer vision.”
Read more about all the projects in the recent round of Australian Research Council (ARC) Future Fellowships: ANU wins $11.8m in ARC Future Fellowships