Infusing Definiteness into Randomness: Rethinking Composition Styles for Deep Image Matting

We study the composition style in deep image matting, a notion that characterizes data generation flow on how to exploit limited foregrounds and random backgrounds form training dataset. Prior art executes this completely manner by simply going through foreground pool or optionally combining two before foreground-background composition. In work, we first show naive combination can be problematic therefore derive an alternative formulation reasonably combine foregrounds. Our second contribution is observation matting performance benefit from certain occurrence frequency of combined their associated source during training. Inspired this, introduce novel binds definite triplet. addition, also find different orders lead patterns, which further inspires quadruplet-based style. Results under controlled experiments four baselines our styles outperform existing ones invite consistent improvement both composited real-world datasets. Code available at: https://github.com/coconuthust/composition_styles