LG AI Research Paper Selected for ICCV 2021

Papers co-authored by Sungha Choi, and Yeonsik Jo from Vision Lab at LG AI Research were selected for ICCV 2021, an international computer vision event. Vision Lab at LG AI Research studies various subjects including Visual analytics, as well as Large-scale vision models, and has continuously shown great achievements following CVPR 2021.



Standardized Max Logit: A Simple yet Effective Approach for Identifying Unexpected Road Obstacles in Urban-scene Segmentation (Link)
Sanghun Jung(KAIST AI), Jungsoo Lee(KAIST AI), Daehoon Gwak(KAIST AI), Sungha Choi, Jaegul Choo(KAIST AI)
During model training for a semantic segmentation task, it enables the system to model uncertainties of new objects, that were not included in the training, and classify them into an ‘unknown’ category rather than an error category. In other words, it recognizes obstacles on the road as ‘unknown’ objects, and not as part of the road. Therefore, it is very useful for safety-critical application such as autonomous driving.


Rethinking Deep Image Prior for Denoising (Link)
Yeonsik Jo, Se Young Chun(Seoul national university), Jonghyun Choi(GIST)
It re-examines the operation principle of the preexisting ‘Deep image prior’ with effective degrees of freedom. Based on such analysis, it suggests an algorithm that removes noise only by using one image without any training data.


*ICCV is the most prestigious international scholastic society within the field of computer vision and pattern recognition along with CVPR. It is an international conference where global research institutes and IT companies share research outcomes. ICCV 2021 will be held online from October 11 to October 17. (http://iccv2021.thecvf.com/home)