CVPR Daily - Thursday

DAILY T h u r s d a y Poster Presentation 16 Yan Wang is a third-year PhD student at Cornell University working on computer vision and self-driving cars. Wei-Lun “Harry” Chao is a first- year assistant professor at Ohio State University working on machine learning and computer vision. He and Yan met while Harry was a postdoc at Cornell. Their work is about finding a solution to a 3D object detection adaptation challenge in autonomous driving. They speak to us ahead of their poster presentation today (Thursday). Self-driving cars are the subject of many papers and there are extraordinary global efforts to develop real-world applications. Deep learning is used for perception and has brought some substantial improvements to the field, but deep networks are very easy to overfit to some specific cases, including brightness, car size and models that may appear throughout the data. When people collect self-driving data, they often do it in a small city within one country and under similar weather conditions. When models are trained on that dataset, it’s easy to overfit to these conditions, and it cannot generalize well on other datasets. To solve this, Yan and Harry have performed extensive analysis to see which factors cause the performance to drop. What they found was that the problem was largely due to varying car size in different geographic regions. They have designed a simple method based on the average car size to yield a strong correction of the adaptation gap. Harry tells us they use a 3D object detector called PointRCNN. This is an algorithm which looks at a point cloud of the 3D environment to find out which part of the point cloud is a car. It then outputs the car size, rotation and location. The problem is that in different geolocations you may see the point cloud become a little larger. For example, in the US, the most common car is the Ford F-series. They are 5-meter long trucks. Whereas, in Germany, the most popular car is a 4-meter long Volkswagen Golf. Train in Germany, Test in the USA: Making 3D Object Detectors Generalize Yan Wang