Sketch

“我们提供了一个新的配对图像和轮廓图数据集，用于研究视觉理解和草图生成。在这个数据集中，有 1,000 张户外图像，每张图像与 5 幅人类图画配对（总共 5,000 张图画）。绘图的笔画与图像边界大致对齐，更容易将人的笔画与图像边缘对应起来。数据集是使用 Amazon Mechanical Turk 收集的。要求 Turker 追踪模糊的背景图像。为了获得高质量的注释，我们设计了一个带有详细说明页面的标签界面，包括许多正面和负面的例子。通过将以下类型的注释视为拒绝候选，通过人工检查实现质量控制：（1）缺少内边界，（2）缺少重要对象， (3) 与原始边缘有很大的偏差，(4) 内容无法识别，(5) 用简笔画画人，(6) 在空白区域有阴影。因此，在广告中除了接受的 5,000 份图纸，我们还有 1,947 份被拒绝的提交，可用于设置自动质量守卫。”

We present a novel paired image and sketch dataset for research on visual understanding and sketch generation. This dataset includes 1,000 outdoor images, with each image paired with 5 human-drawn sketches, totaling 5,000 sketches. The sketch strokes are roughly aligned with the image boundaries, which facilitates matching human-drawn strokes to the edges of the original images. The dataset was collected using Amazon Mechanical Turk. Workers (Turkers) were instructed to trace over blurry background images. To ensure high-quality annotations, we designed a labeling interface equipped with a detailed instruction page, including numerous positive and negative examples. Quality control was conducted via manual inspection, where submissions were rejected if they fell into the following categories: (1) missing internal boundaries, (2) missing important objects, (3) exhibiting large deviations from the original edges, (4) having unrecognizable content, (5) featuring stick-figure drawings of humans, (6) having shading in blank areas. Thus, in addition to the 5,000 accepted sketches, we have 1,947 rejected submissions that can be utilized to develop automated quality guardrails.

搜集汇总

背景与挑战

背景概述

Sketch数据集是一个包含1,000张户外图像和5,000幅对应轮廓图的配对数据集，用于视觉理解和草图生成研究。数据集通过Amazon Mechanical Turk收集，并经过严格的质量控制，确保绘图与图像边缘对齐且内容可识别。