Computer Vision News - February 2023

4 Best Paper Award at IEEE ISM “ Explanation methods are extremely important for real-world applications, ” Nikolaos tells us. “ Until now, classifiers have been black boxes. The user sees the result of the classifier and can’t understand why it has taken a decision. Usually, they don’t know about deep learning and are not computer scientists. They just need an explanation for why the classifier has correctly recognized or failed to TAME is a CAM-based method for explaining the decisions of a convolutional neural network classifier . It builds on previous explanation methods for classifiers in a relatively new and underexplored field, where there’s been plenty of focus on improving the accuracy of classifiers but less on explaining why they reach their decisions. Vasileios Mezaris is a Research Director and Head of the Intelligent Digital Transformation Laboratory (IDT Lab) at CERTH/ITI, where Nikolaos Gkalelis is a senior postdoctoral researcher specializing in deep learning for multimedia, video, and image analysis. Together with first author Mariano Ntrougkas, they recently won the Best Paper Award at the International Symposium on Multimedia (ISM 2022). Vasileios and Nikolaos are here to talk to us about their award-winning work. TAME: ATTENTION MECHANISM BASED FEATURE FUSION FOR GENERATING EXPLANATION MAPS OF CONVOLUTIONAL NEURAL NETWORKS BEST PAPER AWARD IEEE ISM Mariano Ntrougkas Nikolaos Gkalelis Vasileios Mezaris