CVPR Daily - Thursday

DAILY T h u r s d a y 17 Yan Wang “As humans, when we look at a car coming towards us, we only see the front, but we can imagine how long the car is because of our memory of seeing similar vehicles,” Harry explains. “When we visit different countries, we can’t apply the same memory. This solution involves the machine learning that a car which may be 5 meters long in one environment, could be 4 meters long in a new environment. We adjust the data and retrain our model by changing the size of the point cloud in training to make the original environment and new environment look very similar. By doing this, the model can be applied in different settings and universally fit.” In terms of next steps for the work, Harry sees it providing a strong baseline for formulating 3D object detection into a domain adaptation problem. He believes other researchers could utilize the way they collect, formulate, and unify the data, and they plan to build a new algorithm to support this. Looking to the future, he would like to see car companies use this work as a basis for generalizing their systems to different environments, and perhaps even share data to cover more of the world. The overall aim always being to make the system more robust and get closer to the end goal of seeing self-driving cars becoming a part of our everyday life. Yan tells us what he has found most exciting about this work: “Previously Wei-Lun Chao