Clothing Dataset
5,000+ Images of Clothes in Public Domain
Two months ago I asked the community to help me collect a dataset with clothes. And now it’s ready!
You can download it here:
- Full size: https://www.kaggle.com/agrigorev/clothing-dataset-full
- Resized: https://github.com/alexeygrigorev/clothing-dataset
- A subset with 10 most popular classes, resized: https://github.com/alexeygrigorev/clothing-dataset-small
In total, we have collected 5,000 images of 20 types of clothes. The dataset is released under a public domain license (CC0). This means that anyone can use this data for any purpose, also commercial.
For example:
- Training a model for self-education
- Creating a tutorial or a course (free or paid)
- Writing a book
In this post, we’ll tell you more about the dataset. In particular, we’ll cover:
- Different ways of collecting data: networking, Yandex.Toloka, and Tagias
- Labeling the data
- Accessing and using the dataset
Let’s start!
Dataset collection
We used three different ways to collect the dataset:
- Toloka — a crowdsourcing platform
- A networking crowdsourcing initiative on social media
- Tagias — a company specializing in data collection
Toloka
1,400 images (28% of the entire dataset) were collected using Toloka — a crowdsourcing platform. The task was simple: take a picture of some clothes and upload it.
We also experimented with Amazon Mechanical Turk, but creating tasks with image uploads is a lot more difficult on Amazon. In Toloka, setting up an image collection form requires no programming at all.
The results from Toloka contained:
- Valid images that actually were taken by the workers
- Images of clothes downloaded from the Internet
- Images of random objects, not clothes
- A lot of duplicates
To check if an image was downloaded from the Internet, we used Google.
Collecting data from Toloka required careful analysis of the results and in the end, we discarded more than 40% of the submitted images. That’s why we decided to try other options.
Network crowdsourcing
One option was to use the power of social media. We asked the network on LinkedIn and Twitter to contribute to the dataset. To do it, we created a call-for-action article and shared it.
To collect the data, we set up two simple Airtable forms. We explicitly asked the participants to confirm that the images are their own and they are willing to share them under CC0.
To give additional motivation to upload images, we gave away 3 copies of Machine Learning Bookcamp to the top 3 contributors. In the end, 32 people submitted their images, with 600 images in total.
Tagias
The campaign on social media got a lot of attention and we got contacted by multiple companies. One of them, Tagias, agreed to contribute to the project and tremendously helped with data collection.
Workers at Tagias take pictures themselves and don’t scrape them from the Internet. There’s an internal validation process, so we didn’t need to verify that the images are genuine. Only a couple of provided images weren’t suitable, and our feedback was taken into account.
In total, Tagias contributed 3000 images, which is 60% of the entire dataset.
Labeling
After collecting enough images, we needed to label them. We did it manually without crowdsourcing using a simple annotation tool based on IPython widgets.
While labeling, we sometimes made mistakes. To correct these mistakes, we used a simple idea based on training a neural network:
- First, train a model with a high learning rate for a couple of epochs on all the data.
- Next, apply the model to the training data and check where it makes mistakes.
- If the network is correct and there’s a mistake in data, correct it.
Accessing the data
Full dataset
There are multiple ways to download the dataset:
- Original images: https://www.kaggle.com/agrigorev/clothing-dataset-full
- Resized images: https://github.com/alexeygrigorev/clothing-dataset
The dataset contains 20 classes:
- T-Shirt (1011 items)
- Long Sleeve (699 items)
- Pants (692 items)
- Shoes (431 items)
- Shirt (378 items)
- Dress (357 items)
- Outwear (312 items)
- Shorts (308 items)
- Hat (171 items)
- Skirt (155 items)
- Polo (120 items)
- Undershirt (118 items)
- Blazer (109 items)
- Hoodie (100 items)
- Body (69 items)
- Top (43 items)
- Blouse (23 items)
We marked images of children clothes with a special flag “kids”:
- True (476 items)
- False (4927 items)
Some of the items are still labeled “Not sure”, “Others”, or “Skip”, and there could be labeling errors. The corrections are welcome. The best way to submit a correction is to create a pull request via GitHub: https://github.com/alexeygrigorev/clothing-dataset
Top-10 subset
Images of some classes don’t appear very often in the dataset. This means that training a model on this data is difficult — to train a meaningful model, we need at least 100–200 images of each class.
That’s why, for educational purposes, we created a subset of the full dataset that covers only the top-10 classes:
- T-shirt (928 items)
- Long Sleeve (576 items)
- Pants (559 items)
- Shirt (345 items)
- Shoes (297 items)
- Dress (288 items)
- Shorts (257 items)
- Outwear (246 items)
- Hat (149 items)
- Skirt (136 items)
For this dataset, we excluded images of clothes for children.
To make it simpler, we already resized the images, so it’s faster to load them while training a model. The dataset is already split into train, validation and test.
You can download it from https://github.com/alexeygrigorev/clothing-dataset-small.
This dataset was used in Chapter 7 of Machine Learning Bookcamp. You can check a code example in the book’s GitHub repository.
Now it’s your turn! Tell us about your ideas how you’d like to use it!
Summary
- We collected over 5,000 images and for that, we relied on three sources: Tagias, Toloka, and networking
- Submissions from Toloka required a lot of extra analysis
- Labeling was done manually using IPython widgets and we corrected labeling mistakes using a simple neural network.
- The dataset is available in two parts: a full dataset and a subset (top-10 classes).
Acknowledgments
We’d like to thank
- Kenes Shangereyev and Tagias.com for helping with 3000 images
- All the 32 people who contributed their images to the dataset via the forms
- Everyone who supported the initiative by engaging with the announcements on social media
