This is a edge-map dataset including 290281 edge-maps corresponding to 345 object categories, termed as Edge-Map-345C, established by Peng Xu. In particular, these 345 categories are corresponding to the 345 free-hand sketch categories of Google QuickDraw dataset. Please see categories.txt for details.
Some edge-mage samples are shown in above figure. This repository provides zip format packages in Google Drive and Baiduyun Disk, and will be continuously updated! Our paper associated with this dataset will be released soon. If this dataset can be used in your research, please cite our paper.
For sketch-specific study, this edge-map dataset can be used to pretrain the networks to solve the training data scarcity issue.
The data collection can be divided into two steps: photo collection and edge-map extraction.
There are 157 (out of 345) classes of QuickDraw having at least one corresponding class in ImageNet. For each of these 157 categories, I manually choose one most similar category from ImageNet. Since many of ImageNet are multi-object images, we crop photos from the annotated bounding box areas, thus only photos with bounding box annotations provided can be used. This cropping operation can alleviate the domain gap between edge-map and sketch. For the remaining 188 categories, we program crawler to download images from Google Images.
After photo collection, we use Edge Boxes toolbox to extract edge-maps from our collected photos. Each edgemap has been located in the center and takes up about 90% area of the photo. We resized all the edge-maps as 3×224×224. We recruit five volunteers to manually remove the messy edge-maps that can not be recognize by human. Finally, we obtain 290, 281 edge-maps across 345 categories (averagely 841 per category).
These 290, 281 edge-maps were randomly divided into training set (263, 655) and validation set (26, 626).
The main preprocessing scripts will be released in ./code/.
I trained some state-of-the-art CNN networks on Edge-Map-345C training set, and classification results on Edge-Map-345C validation set are reported in following table.
| Model | Validation Accuracy |
|---|---|
| MobileNet V1 | 0.6108 |
| MobileNet V2 | 0.5971 |
| ResNet-18 | 0.5943 |
| ResNet-34 | 0.5931 |
| ResNet-50 | 0.5896 |
| ResNet-152 | 0.5829 |
| DenseNet-121 | 0.6153 |
| DenseNet-161 | 0.5518 |
| DenseNet-169 | 0.5177 |
| DenseNet-201 | 0.5514 |
| GoogLeNet Inception V3 | 0.6237 |
| VGG-11 | 0.6132 |
Please download them via following links, respectively.
Training Set (4.9GB): Baiduyun Disk part 1， Baiduyun Disk part 2; [Google Drive] (coming soon).
Validation Set (490MB): Baiduyun Disk, [Google Drive] (coming soon).
If you would like to download from Baiduyun, Please send email to Peng Xu to get the fetch code.
If these links can not work well, please feel free to contact Peng Xu (peng.xu [AT] ntu.edu.sg).
If you find this dataset useful, please cite our paper as the following bibtex:
@article{xu2020learning,
title={On Learning Semantic Representations for Million-Scale Free-Hand Sketches},
author={Xu, Peng and Huang, Yongye and Yuan, Tongtong and Xiang, Tao and Hospedales, Timothy M and Song, Yi-Zhe and Wang, Liang},
journal={arXiv preprint arXiv:2007.04101},
year={2020}
}