PPR10K: A Large-Scale Portrait Photo Retouching Dataset with Human-Region Mask and Group-Level Consistency (2105.09180v1)

Abstract: Different from general photo retouching tasks, portrait photo retouching (PPR), which aims to enhance the visual quality of a collection of flat-looking portrait photos, has its special and practical requirements such as human-region priority (HRP) and group-level consistency (GLC). HRP requires that more attention should be paid to human regions, while GLC requires that a group of portrait photos should be retouched to a consistent tone. Models trained on existing general photo retouching datasets, however, can hardly meet these requirements of PPR. To facilitate the research on this high-frequency task, we construct a large-scale PPR dataset, namely PPR10K, which is the first of its kind to our best knowledge. PPR10K contains $1, 681$ groups and $11, 161$ high-quality raw portrait photos in total. High-resolution segmentation masks of human regions are provided. Each raw photo is retouched by three experts, while they elaborately adjust each group of photos to have consistent tones. We define a set of objective measures to evaluate the performance of PPR and propose strategies to learn PPR models with good HRP and GLC performance. The constructed PPR10K dataset provides a good benchmark for studying automatic PPR methods, and experiments demonstrate that the proposed learning strategies are effective to improve the retouching performance. Datasets and codes are available: https://github.com/csjliang/PPR10K.

• The paper presents PPR10K, a dataset of 11,161 high-quality raw portrait photos grouped into 1,681 sets with annotated human-region masks.
• It introduces novel evaluation metrics, including human-centered PSNR, CIELAB differences, and Group-Level Consistency (GLC), to assess retouching performance.
• Experimental results demonstrate that models trained on PPR10K achieve superior retouch quality and consistency compared to those using general enhancement datasets.

Insights on "PPR10K: A Large-Scale Portrait Photo Retouching Dataset with Human-Region Mask and Group-Level Consistency"

The presented paper discusses the creation and validation of a new dataset named PPR10K, designed specifically for the task of Portrait Photo Retouching (PPR). PPR involves distinct characteristics from general photo retouching, such as the requirement to prioritize attention on human regions and maintain a consistent tone for a group of photos. The paper elucidates the challenges posed by existing datasets, which inadequately address these specific requirements, thereby motivating the construction of PPR10K.

Dataset Construction and Characteristics

PPR10K is a large-scale dataset consisting of 11,161 high-quality raw portrait photos organized into 1,681 groups. The human-region masks offered by the dataset facilitate improved attention towards subject areas. Each photo has been meticulously retouched by three experts aiming for consistent tones across a group, which is critical given the variations in subject views, lighting, and camera settings. The authors also include high-resolution segmentation masks to assist in focusing on human regions.

Evaluation Measures and Learning Strategies

The paper defines several objective measures pertinent to evaluating PPR: PSNR and CIELAB color difference are utilized similarly to those in general photo enhancement tasks. However, to address the unique elements of PPR, the authors define human-centered versions of these metrics and introduce a Group-Level Consistency (GLC) measure. The GLC measure is specifically designed to quantify tonal consistency across photos in a group using statistics of color components in the CIELAB space.

Moreover, novel learning strategies are proposed to optimize these measures. The first strategy, Human-Region Priority (HRP), weights the loss function to pay more attention to human regions, consequently improving the visual quality in those areas. The second strategy for achieving GLC involves simulating intra-group variations by introducing slight transformations within single images during training, which effectively enhances the model’s robustness to content variation within groups.

Performance and Comparison

The authors validate these strategies using state-of-the-art retouching models such as HDRNet, CSRNet, and 3D LUT. The experimental results demonstrate that models trained on PPR10K outperform those trained on general enhancement datasets, such as FiveK, when evaluated on PPR tasks. This performance is indicated by superior metrics and visual quality, thus underlining the importance of PPR10K’s derivatives like human-region masks and group-level consistency data.

Practical Implications and Future Directions

PPR10K offers a beneficial benchmark for researchers focusing on automatic portrait retouching by providing data that aligns more closely with professional standards and real-world applications. The implications of adopting such a dataset span improved user engagement in photo-centric applications to more ergonomic workflows in professional photography settings.

Looking forward, further research could explore more advanced models leveraging PPR10K for tasks beyond retouching, such as portrait segmentation or stylization, by making use of the rich annotations available. Additionally, addressing challenges related to computational efficiency while preserving high visual fidelity in real-time applications remains a promising avenue for future inquiries. The dataset’s scalability and adaptability also allow for potential applications in training and evaluating generative models aimed at personalization in photographic retouching.