About annotation and dataset managment
In this article I raise the issue of image annotation and dataset management. In deep learning and machine learning applicatiaons we often come across problems not directly related to our algorithm, but to data management.
- How to store your images safely and access them from any location?
- How to precisely control who has access to what?
- How to add image annotations? And how to mark data?
- How to distribute the annotation work between multiple workers?
- How to track progress, and document how annotations have been created?
- How to assure high quality of annotations through a two-stage process of annotation and review?
- How to make your deep learning experiments fully repeatable by referring to easily trackable datasets and annotation projects?
In this article you will find the answers. One of the most popular and advanced solutions is Zillin, and today I will show you how it could be used.
Zillin – online tool for image annotation and dataset management
Zillin.io is a secure online tool for image annotation and dataset management. Adaptive Vision acquired this service in January this year. The new version had been anounced and re-designed version will meet the needs of machine vision market. For Adaptive Vision, it is a very important extension of its existing portfolio of deep learning products. Zillin will complement Adaptive Vision Deep Learning Add-on. You can read more about it in this post.
In the photo: Michał Czardybon (Adaptive Vision, CEO) and Zbigniew Kawalec (QZ Solutions, CEO)
One of the most important features of Zillin is to control permission levels. Depending on the role, user has different possibilites:
- Manager – can add or remove people from the team.
- Developer – defines projects and creates models using annotated images.
- Collaborator – uploads images and makes annotations.
- Guest – can see the projects, but in the read-only way.
- Worker – only gets individual images for annotation in a batch mode.
Moreover, each user can have different role in different projects. There are two roles: Annotator and Reviewer. Depending on the premision level, reviewer can also annotate data. Those functions allows you to:
- Distribute the annotation work between multiple workers.
- Transfer images between your team and your customers.
- Precisely control who has access to what.
- Assure high quality of annotations through a two-stage process of annotation and review.
Due to online structure, you can store your images safely and access them from any location. As you can see, in the image below there are also a lot of other options like publishing, deleting and downloading dataset. You are also able to check dataset status, number of images, total size and other features.
When your dataset is ready, permission levels and roles are assigned, you can start to annotate data. You can create annotations using several different tools. Each tool will have a class name attached.
- Bounding Boxes – simply draw frames to quickly select interesting objects.
- Polygons – create precise outlines of objects or defects.
- Points – mark key points like human eyes, object corners or picking slots.
- Polylines – useful, for example, for marking road lanes or veins.
- Questions – add meta-information which can be Yes/No, numerical or textual.
In the following example I used polygons option to mark different types of nuts. It is also worth mentioning, that several people can work on the annotation simultaneously.
Annotator can now mark objects in each image. In some cases, he can also skip data and choose one of the available options, that justify his decision to skip the image.
After that, reviewer is able to review annotation and accept or reject them. If reviewer have permision he can also edit annotation. In the last step, data can be easly exported, so you can use it for your deep learning application.
Summary and some news
As you could see, Zillin.io is an online tool, which has many uses for annotation and dataset management. In my opinion, this is a helpful tool, if you have to work in the group of several people on one project. The tool is still improving and in the next update we will obtain many useful features like other type of exported data (Adaptive Vision’s Deep Learning fans should be satisfied 😀) and many others facilities for the user.