ICCV Daily 2021 - Wednesday

All the techniques used are based on computer vision techniques . The work uses a semantic segmentation model which shows state-of-the-art performance on segmenting the classes in an image. After that, as a post- processing approach, Gaussian smoothing is applied, which uses a Gaussian kernel for removing the smooth noises in an image. One of the assumptions in the work is that the model has a reasonable capacity to discriminate between in-distribution and out-of-distribution pixels. If the model cannot discriminate between the two in the first place, the rest of it does not work. In terms of next steps, the team want to improve its capability to discriminate between those in-distribution and out-of-distribution pixels before applying the three models. “ Our method achieves state-of-the-art performance on the public leaderboard, ” Sanghun says proudly. “ It’s a very simple and effective approach. It doesn’t need any additional training or additional out-of-distribution data set. ” 20 DAILY ICCV Wednesday “One of the reasons is because the boundary regions are where the classes change, like from cars to road.” Oral Presentation