CVPR Daily - Thursday

DAILY T h u r s d a y Poster Presentation 18 people could only work on one dataset, theKITTI dataset, but right nowthey have five datasets in the same format, so they can easily do experiments using them all to verify the performance of their model. By simply changing the car size we can minimize the domain difference and drastically improve performance. People haven’t tried domain adaptation for 3D object detection before. Now, they have a good dataset to do that. We’re very proud about that.” Harry agrees: “Just like Yan said, a very simple algorithm can improve performance a lot. Nowadays, because machine learning is so powerful, people just throw all their data in and hope the machine can learn. For us though, we did a lot of analysis and found some very simple things. Even if you don’t apply deep learning you can begin to solve your problem, to a certain extent. This is something people should always remember: when you have a new problem, start with a simple solution. Once you have pushed it to its limits, then look for something more sophisticated.” To find out more about Yan and Harry’s work, please visit one of the Poster 3.2 Q&A sessions today (Thursday) at 11:00-13:00 or 23:00-00:00.