Sequential Feature Filtering Classifier (IEEE Access)

Minseok Seo, Jaemin Lee, Jongchan Park, Daehan Kim and Dong-Geol

Abstract

In this study, we propose Sequential Feature Filtering Classifier (FFC), a simple but effective classifier for convolutional neural networks (CNNs). Using sequential LayerNorm and ReLU, FFC zeroes out low-activation units and preserves high-activation units. The sequential feature filtering process generates multiple features, which are transmitted to a shared classifier, yielding multiple outputs. FFC can be applied to any CNN with a classifier and it significantly improves the performance with negligible overhead. In this study, the efficacy of FFC is validated extensively on various tasks—ImageNet-1K classification, MS COCO detection, Cityscapes segmentation, and HMDB51 action recognition. Moreover, it is empirically established that FFC can further improve performances using additional techniques, including attention modules.

Datasets

The datsets can be downloaded from

ImageNet-1K

/IMAGENET/ILSVRC/Data/CLS-LOC/

Directory tree

   DATA/
       ILSVRC/ 
             /Data
                   /CLS-LOC
                           /train
                           /val
       
   Sequential-Feature-Filtering-Classifier/
       (main.py)

Models

Note that our top performance was 76.97 as a result of multiple runs, but the average of the results of multiple runs was 76.84 on the paper.

METHOD	DATASET	ACC
ResNet-50	ImageNet-1K	75.80
ResNet-50+FFC	ImageNet-1K	76.97

train-test script

train

python main.py --ffc

test

python main.py --ffc --r