SketchHairSalon

SketchHairSalon，一个两阶段框架，用于直接从描绘所需头发结构和外观的手绘草图生成逼真的头发图像。在第一阶段，我们训练一个网络从输入的头发草图中预测头发遮罩，并带有一组可选的非头发笔触。在第二阶段，训练另一个网络从输入草图和生成的遮罩中合成头发图像的结构和外观。为了使两个阶段的网络意识到笔画的长期依赖性，我们对它们应用了自我注意模块。为了训练这些网络，我们提出了一个新的数据集，其中包含数千个带注释的头发草图图像对和相应的头发遮罩。

SketchHairSalon is a two-stage framework for generating photorealistic hair images directly from hand-drawn sketches that depict the desired structure and appearance of the target hair. In the first stage, we train a neural network to predict hair masks from input hair sketches along with a set of optional non-hair strokes. In the second stage, another neural network is trained to synthesize the structure and appearance of hair images from the input sketches and the generated masks. To enable the two-stage networks to capture long-range dependencies of the strokes, we apply self-attention modules to both networks. To train these networks, we propose a novel dataset containing thousands of annotated paired hair sketch images and their corresponding hair masks.

搜集汇总

背景与挑战

背景概述

SketchHairSalon是一个用于从手绘草图生成逼真头发图像的两阶段框架，通过预测头发遮罩和合成图像实现，并应用了自我注意模块。该数据集由香港城市大学、浙江大学和香港中文大学于2021年发布，用于训练相关网络。