VizWiz-FewShot

我们介绍了一些拍摄的本地化数据集，这些数据集来自真正试图了解他们拍摄的图像中的视觉内容的摄影师。它包括由视觉障碍者拍摄的4,500多个图像中100类别的近10,000个分割。与现有的少拍摄对象检测和实例分割数据集相比，我们的数据集是第一个定位对象中的孔 (例如，在我们的12.3% 分割中发现)，它显示了相对于图像占用更大范围的对象，文本在我们的对象中是常见的五倍以上 (例如，在我们的22.4% 分割中找到)。

We present a curated local dataset collected by photographers who genuinely strive to comprehend the visual content of the images they captured. The dataset contains nearly 10,000 segmentations across 100 categories, sourced from over 4,500 images taken by visually impaired photographers. Compared with existing few-shot object detection and instance segmentation datasets, our dataset is the first to localize holes within segmented objects (e.g., observed in 12.3% of our segmentations). Additionally, our dataset demonstrates that for objects that occupy a larger relative proportion of the image, text appears in these objects more than five times as frequently as in other objects, with text detected in 22.4% of our segmentations.

搜集汇总

背景与挑战

背景概述

VizWiz-FewShot是一个包含4,500多张图像的数据集，由视觉障碍者拍摄，涵盖100个类别的近10,000个分割实例。该数据集首次在少样本目标检测和实例分割中定位对象中的孔，并具有对象占比较大和文本出现频率较高的特点。