Sequential Feature Filtering Classifier

The Ensemble and mixture of expertise method is the most intuitive simple way to improve performance in field recognition using convolutional neural networks (CNNs). However, difficult apply real-time operation applications because amount computational overhead parameters increase proportion number models. In another, a that extracts various combines it cumbersome requires large change network. this study, we propose Sequential Feature Filtering Classifier (FFC), but effective classifier for CNNs. Using sequential LayerNorm ReLU, FFC zeroes out low-activation units preserves high-activation units. feature filtering process generates multiple features, which are transmitted shared classifier, yielding outputs (multiple expertise). can be applied any CNN with significantly improves negligible overhead. efficacy validated extensively on tasks—ImageNet-1K classification, MS COCO detection, Cityscapes segmentation, HMDB51 action recognition. Moreover, empirically established further performances additional techniques, including attention modules. Code available at https://github.com/seominseok0429/Sequential-Feature-Filtering-Classifier .