CVPR Daily - Tuesday

Regarding the computer vision techniques involved in this work, Jiawei identifies that the aim was to maximize both the inter-class separation and the intra-class compactness , also called the feature distribution’s geometry. “ For maximizing the inter-class separation, we use a classifier called simplex equiangular tight frame , ” he explains. “ Simplex ETF is offline- derived and fixed during the training. One property of this classifier is that the weights are all maximally and equally separated in the entire feature space. From the perspective of maximizing intra-class compactness, many different approaches can be used, but we use the simplest one to highlight the importance of this conceptual idea. We do that by adding margins during the training. For conventional cross-entropy loss, we directly minimize the difference between the probability distribution predicted by the model and the ground-truth vector. We add margins in the conventional cross-entropy loss to force the features close to the corresponding class center. ” 14 DAILY CVPR Tuesday Poster Presentation