Last week we presented our most recent works at CARS conference 2017. First, Quentin Angermann presented our work done in collaboration with ENSEA and Hopital de Saint Antoine in order to build up a real-time polyp detection method. In this case, our work proposes an adaptation of already existing still-frame based polyp detection methods to make them work when dealing with video sequences. The methodology is tested on a completely new fully annotated video database (available at https://grand-challenge.org/site/endovissub2017-giana), showing promising results with respect to the number of polyps detected and computational requirements.
Second, Jorge Bernal presented his work (done in collaboration with David Vázquez from CVC and researchers from MILA Montreal) on semantic segmentation on the whole endoluminal scene. This work aimed to present a new benchmark on endoluminal scene segmentation. Results of a comparative study between classic hand-crafted approaches and trending techniques such as CNNs show that the latter are the new state-of-the-art, outperforming classic approaches by a large margin with a reduced processing time.
We will publish some slides soon so stay tuned!