Reproduce “The Lovász-Softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Maxim Berman, Amal Rannen Triki, Matthew B. Blaschko
Researchers from KU Leuven introduced the Lovász-Softmax loss, a differentiable surrogate for the Jaccard index (IoU) measure, enabling direct optimization of this standard evaluation metric in semantic image segmentation. This method achieved mIoU improvements of 3-4% on datasets like Pascal VOC and Cityscapes compared to traditional cross-entropy loss.
Key insights
- The Jaccard loss can be expressed as a submodular function, enabling the use of its Lovász extension to derive a continuous, differentiable, and convex surrogate for optimization.
- Aligning the training objective directly with the evaluation metric (IoU) through a mathematically sound surrogate yields substantial and consistent improvements in segmentation quality.
- The proposed loss naturally produces more coherent and perceptually accurate segmentations by better handling object boundaries and small instances.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.