Core Concepts
The author proposes an End-to-End Human Instance Matting framework to efficiently estimate alpha mattes for multiple human instances simultaneously.
Abstract
The paper introduces the E2E-HIM framework for human instance matting, addressing challenges in accuracy and computational efficiency. It outperforms existing methods on various datasets, showcasing significant improvements in speed and error rates.
Stats
Experiments on HIM-100K demonstrate 50% lower errors and 5× faster speed.
E2E-HIM achieves competitive performance on traditional human matting datasets.
The proposed method consumes 305 GFlops to estimate all instance-level alpha mattes in a 640 × 640 image.