Reproduce “Constrained-CNN losses for weakly supervised segmentation” on OpenResearch

About this paper

Hoel Kervadec, Jose Dolz, Meng Tang, Eric Granger, Yuri Boykov, Ismail Ben Ayed

This work introduces a differentiable penalty incorporated directly into the loss function of deep convolutional neural networks, enabling the effective use of global inequality constraints for weakly supervised semantic segmentation. It achieves segmentation performance approaching fully supervised methods, particularly with individual region size bounds, while maintaining computational efficiency comparable to standard training.

Key insights

• Directly incorporating global inequality constraints as a differentiable penalty is computationally efficient and more effective than traditional Lagrangian dual optimization or proposal-based methods for weakly supervised segmentation.
• More precise constraint information, such as individual bounds for each image, significantly narrows the performance gap between weakly and fully supervised learning.
• The approach offers consistent performance gains across various medical imaging datasets and weak annotation strategies, highlighting its robustness and generalizability.

Summary by alphaXiv.