Reproduce “Boosting Robustness of Image Matting with Context Assembling and Strong Data Augmentation” on OpenResearch

About this paper

Yutong Dai, Brian Price, He Zhang, Chunhua Shen

The RMat framework from researchers at The University of Adelaide, Adobe Inc., and Zhejiang University enhances image matting robustness and generalization by integrating a hybrid CNN-Transformer architecture with strong data augmentation techniques. This approach improves trimap robustness and real-world performance without sacrificing benchmark accuracy, achieving an approximate 11% improvement on SAD and 27% on Grad metrics on the Composition-1k dataset.

Key insights

• The hybrid CNN-Transformer architecture effectively captures both global contextual understanding and fine-grained local details, leading to a more comprehensive representation for matting.
• Previous data augmentation methods for matting often failed to generalize without trade-offs because they violated the linear matting equation, especially for region-wise or nonlinear pixel-wise transformations applied to the composited image.
• Generating pseudo-labels for augmented samples that violate the matting equation allows the network to learn from augmented data without relying on incorrect ground truths, thus effectively bridging the domain gap.

Summary by alphaXiv.