CVPR Daily - Tuesday

“ Then, when you try to transform back from the discrete coordinate to the continuous coordinate, you need to consider this rounding. We added a 0.5 offset because the discrete coordinates are integer, but the continuous coordinate could be an arbitrary floating-point number. We always add a 0.5 offset to the integer when we transform back to make up this rounding. ” We ask Bowen what he thinks convinced the area and program chairs that his paper was good enough for an oral this year, and he points toward its surprising results . “We’re not the first to try to reduce annotation time for instance segmentation, but in the past, when people tried to use weak supervision with less cost than the mask annotation, they always traded accuracy for annotation time, ” he points out. “ Even the best methods can only achieve around 80% of the fully supervised method, whereas we get as close as 98%. ” Even with a figure so close to 100%, Bowen says that if people aim for accuracy over reduced annotation time, they will still opt for mask annotation. Therefore, in terms of the next steps, he is curious to know if they could achieve 100% with much cheaper annotation . Bowen completed this work whilst interning at Facebook AI Research (now Meta), and he is keen to thank his mentor Alexander Kirillov, the last author of this work, for all his support: “ Alex is an incredible mentor. He always gives me insightful feedback, from discussing high-level ideas to digging into low-level technical or implementation details. This paper wouldn’t be as successful without his support! ” To learn more about Bowen’s work [ID 44], come to oral session 1.2.1 today at 13:30 and poster session 1.2. 5 DAILY CVPR Tuesday Bowen Cheng