Contour Drawing Dataset

我们提出了一个新的成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，每张都与 5 幅人类图画配对（总共 5,000 张图画）。这些绘图的笔触与图像边界大致对齐，从而更容易将人的笔触与图像边缘对应起来。

数据集是使用 Amazon Mechanical Turk 收集的。土耳其人被要求追踪昏暗的背景图像。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注界面，其中包含许多正面和负面的例子。质量控制是通过人工检查来实现的，将以下类型的注释视为拒绝候选：（1）缺少内部边界，（2）缺少重要对象，（3）与原始边缘有很大的错位，（4）内容无法识别, (5) 用简笔画画人, (6) 在空白区域涂上阴影。因此，除了接受的 5,000 张图纸外，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 introduce a novel paired image and contour map dataset for visual understanding and sketch generation research. This dataset contains 1,000 outdoor images, each paired with 5 human-drawn sketches, totaling 5,000 sketches. The strokes of these sketches are roughly aligned with the image boundaries, making it easier to map human-drawn strokes to image edges. The dataset was collected via Amazon Mechanical Turk. Annotators participating in the platform were instructed to trace dimmed background images. To obtain high-quality annotations, we designed an annotation interface with a detailed instruction page that includes numerous positive and negative examples. Quality control was conducted through manual inspection, with annotations belonging to the following categories marked as rejected candidates: (1) lack of internal boundaries; (2) omission of important objects; (3) significant misalignment with original edges; (4) unrecognizable content; (5) depicting humans with stick figures; (6) applying shading to blank areas. Thus, aside from the 5,000 accepted sketches, we also have 1,947 rejected submissions that can be utilized to establish automatic quality guardrails.

搜集汇总

背景与挑战

背景概述

Contour Drawing Dataset是一个包含1,000张户外图像和对应5,000幅人类绘制轮廓图的数据集，通过Amazon Mechanical Turk收集并经过严格质量控制，适用于视觉理解和草图生成研究。