tumuyan2/ImageNet-Sketch-HQ

A Selection From https://huggingface.co/datasets/songweig/imagenet_sketch (actually I download from https://opendatalab.com/OpenDataLab/ImageNet-Sketch and the hash is same)

The original ImageNet-Sketch data set consists of 50000 images, 50 images for each of the 1000 ImageNet classes. We construct the data set with Google Image queries "sketch of __", where __ is the standard class name. We only search within the "black and white" color scheme. We initially query 100 images for every class, and then manually clean the pulled images by deleting the irrelevant images and images that are for similar but different classes. For some classes, there are less than 50 images after manually cleaning, and then we augment the data set by flipping and rotating the images.

This dataset is a curated version of ImageNet-Sketch (less than 25000 images), retaining the original file structure and filenames, remove following images

1. Resolution lower than 360x360.
2. Images that were flipped and rotated.

I‘ working on image annotation to remove following images: (It needs long time. If you're interested, you can join in the project. I'm using https://huggingface.co/spaces/tumuyan2/Argilla , with resize & compressed image from https://huggingface.co/datasets/tumuyan2/ImageNet-Sketch-HQ-Resized, result temporarily stored at https://huggingface.co/datasets/tumuyan2/image_net-sketch-hq-resized ):

1. Images with watermarks.
2. Others Low-quality images.

---

中文说明

ImageNet-Sketch (实际上我从 https://opendatalab.com/OpenDataLab/ImageNet-Sketch 下载的，压缩包hash相同)
原数据集由 50000 张图像组成，1000 个 ImageNet 类中的每一个都有 50 张图像。数据集是用谷歌图片查询“sketch of”构建的，其中是标准类名。仅在“黑白”配色方案内进行搜索。每个类最初查询 100 张图像，通过删除不相关的图像和相似但不同类的图像来清理提取的图像。对于某些类，手动清理后的图像少于 50 张，然后通过翻转和旋转图像来扩充数据集。 本数据集对 ImageNet-Sketch 进行精选(数量小于25000)，保留了原有文件结构和文件名，去除了以下图像：

1. 分辨率低于360*360
2. 删除翻转和旋转图像 在进一步筛选剔除以下图像，阶段性成果在 https://huggingface.co/datasets/tumuyan2/ImageNet-Sketch-HQ2 (非常慢，如果你有兴趣可以一起来标注，目前在使用 https://huggingface.co/spaces/tumuyan2/Argilla 标注，使用了压缩分辨率的数据 https://huggingface.co/datasets/tumuyan2/ImageNet-Sketch-HQ-Resized ，标注结果存储在 https://huggingface.co/datasets/tumuyan2/image_net-sketch-hq-resized )：
3. 含水印的图片
4. 低质量（如明显使用nearest放大，非常潦草的简笔画） 待进一步优化图像：
5. 裁切文件边缘