We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.CV

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Computer Vision and Pattern Recognition

Title: Bootstrapping Semantic Segmentation with Regional Contrast

Abstract: We present ReCo, a contrastive learning framework designed at a regional level to assist learning in semantic segmentation. ReCo performs semi-supervised or supervised pixel-level contrastive learning on a sparse set of hard negative pixels, with minimal additional memory footprint. ReCo is easy to implement, being built on top of off-the-shelf segmentation networks, and consistently improves performance in both semi-supervised and supervised semantic segmentation methods, achieving smoother segmentation boundaries and faster convergence. The strongest effect is in semi-supervised learning with very few labels. With ReCo, we achieve 50% mIoU in the CityScapes dataset, whilst requiring only 20 labelled images, improving by 10% relative to the previous state-of-the-art. Code is available at this https URL
Comments: Project Page: this https URL Code: this https URL
Subjects: Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
Cite as: arXiv:2104.04465 [cs.CV]
  (or arXiv:2104.04465v3 [cs.CV] for this version)

Submission history

From: Shikun Liu [view email]
[v1] Fri, 9 Apr 2021 16:26:29 GMT (47644kb,D)
[v2] Mon, 12 Apr 2021 00:53:22 GMT (47644kb,D)
[v3] Wed, 12 May 2021 18:01:39 GMT (47645kb,D)

Link back to: arXiv, form interface, contact.