Основні поняття
Patch-based normalization network for portrait harmonization achieves state-of-the-art results.
Анотація
1. Introduction
Image composition merges elements from different sources.
Realistic composite images require manual adjustments.
Image harmonization aims for visual consistency.
2. Related Work
Traditional color space methods for harmonization.
Recent advances in neural networks for image harmonization.
3. Datasets
Existing datasets for image harmonization.
Introduction of FFHQH dataset for portrait harmonization.
4. Architecture
Patch-based normalization block for global and local statistics.
Patch-based feature extraction module for high-level background representation.
5. Experiments
Training details and stages.
Ablation study on different network configurations.
Evaluation results on iHarmony4 and FFHQH datasets.
6. Evaluation results
Quantitative comparisons with state-of-the-art methods.
Qualitative comparisons showcasing visual coherence.
Limitations in handling non-portrait images and small foreground ratios.
Статистика
최근 방법들은 인코더-디코더 네트워크 [7,8,16] 및 트랜스포머 [9]를 사용하여 조명을 풍부하게 하려고 시도하며, 보조 기능을 이용하여 저수준 표현을 풍부하게 한다.
우리의 최고 설정 모델(D4,5,6 + PFE)은 39.9백만 개의 매개변수를 가지며 디스크 공간을 153MB 소비한다.
PHNet은 인텔 H470 CPU 스레드에서 1.01 FPS, NVIDIA Tesla V100에서 34.49를 달성한다.
Цитати
"Our network achieves state-of-the-art results on the iHarmony4 dataset."
"Our method excels in producing better results at intermediate resolutions."