ICCV Daily 2021 - Wednesday

Semantic segmentation of urban-scene images usually works with a pre- trained model that predicts all the pixels in an image as one of the predefined classes, such as cars or roads. However, there is no training data for out-of-distribution detection . Even if such data could be collected, it would always be restricted because you cannot collect data for every unexpected object in the real world. By detecting anomaly pixels or objects in an image , this work provides an initial point to treat them differently from the training classes. “ There are well-known metrics to measure anomaly scores of the objects in an image, ” Sanghun tells us. 18 DAILY ICCV Wednesday Oral Presentation StandardizedMax Logits: A Simple yet Effective Approach for Identifying Unexpected Road Obstacles in Urban-Scene Segmentation Sanghun Jung (left) and Jungsoo Lee (right) are Master’s students at KAIST AI in South Korea, under the supervision of Professor Jaegul Choo. Their work, which tackles out-of-distribution detection in semantic segmentation, has been selected as an oral presentation this year. They speak to us ahead of their live Q&A session today.