eduzhai > Applied Sciences > Engineering >

Multi-Task Pruning for Semantic Segmentation Networks

  • king
  • (0) Download
  • 20210506
  • Save

... pages left unread,continue reading

Document pages: 15 pages

Abstract: This paper focuses on channel pruning for semantic segmentation networks.There are a large number of works to compress and accelerate deep neuralnetworks in the classification task (e.g., ResNet-50 on ImageNet), but theycannot be straightforwardly applied to the semantic segmentation network thatinvolves an implicit multi-task learning problem. To boost the segmentationperformance, the backbone of semantic segmentation network is often pre-trainedon a large scale classification dataset (e.g., ImageNet), and then optimized onthe desired segmentation dataset. Hence to identify the redundancy insegmentation networks, we present a multi-task channel pruning approach. Theimportance of each convolution filter w.r.t the channel of an arbitrary layerwill be simultaneously determined by the classification and segmentation tasks.In addition, we develop an alternative scheme for optimizing importance scoresof filters in the entire network. Experimental results on several benchmarksillustrate the superiority of the proposed algorithm over the state-of-the-artpruning methods. Notably, we can obtain an about $2 times$ FLOPs reduction onDeepLabv3 with only an about $1 $ mIoU drop on the PASCAL VOC 2012 dataset andan about $1.3 $ mIoU drop on Cityscapes dataset, respectively.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×