eduzhai > Applied Sciences > Engineering >

Mining Cross-Image Semantics for Weakly Supervised Semantic Segmentation

  • king
  • (0) Download
  • 20210505
  • Save

... pages left unread,continue reading

Document pages: 20 pages

Abstract: This paper studies the problem of learning semantic segmentation fromimage-level supervision only. Current popular solutions leverage objectlocalization maps from classifiers as supervision signals, and struggle to makethe localization maps capture more complete object content. Rather thanprevious efforts that primarily focus on intra-image information, we addressthe value of cross-image semantic relations for comprehensive object patternmining. To achieve this, two neural co-attentions are incorporated into theclassifier to complimentarily capture cross-image semantic similarities anddifferences. In particular, given a pair of training images, one co-attentionenforces the classifier to recognize the common semantics from co-attentiveobjects, while the other one, called contrastive co-attention, drives theclassifier to identify the unshared semantics from the rest, uncommon objects.This helps the classifier discover more object patterns and better groundsemantics in image regions. In addition to boosting object pattern learning,the co-attention can leverage context from other related images to improvelocalization map inference, hence eventually benefiting semantic segmentationlearning. More essentially, our algorithm provides a unified framework thathandles well different WSSS settings, i.e., learning WSSS with (1) preciseimage-level supervision only, (2) extra simple single-label data, and (3) extranoisy web data. It sets new state-of-the-arts on all these settings,demonstrating well its efficacy and generalizability. Moreover, our approachranked 1st place in the Weakly-Supervised Semantic Segmentation Track ofCVPR2020 Learning from Imperfect Data Challenge.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×