What is IntelliBrush?
IntelliBrush is Datature's AI-guided labelling tool that allows you to label complex images easily by clicking on the object. It allows the user to make pixel-accurate annotations within a few clicks and does not require pre-training of any models. In essence, it works right out of the box for annotating a wide variety of use cases from blood cells in pathology scans to defects in industrial manufacturing scans.
Traditional Guided Segmentation Tools
From photoshop to labelling tools, developers have been coming up with tools to help users annotate data or select objects faster. The use of foreground segmentation and extraction techniques has been in the scene for some time and we have noticed many entrants in the image labelling space using algorithms within this family — one such famous tool is Deep Extreme Cut or DEXTR.
Users will typically use a range of selection tools or box-drawing mechanism to guide the algorithms on extracting the intended foreground with varying results. More interestingly, we see the use of superpixels algorithms (more on this here) that provides results as such:
Superpixel algorithms are used in various online annotators. This algorithm essentially breaks the images into bounded regions, where users can subsequently select regions that include their objects of interest as a form of generating mask labels. These selected fragments are subsequently "stitched" together either using geometry boolean operations or a concave hull.
However, most of these tools have weaknesses ranging from the speed of prediction, accuracy, number of clicks required, and most importantly, the amount of control for refining your selection. For some of these algorithms, the results are sensitive to image noise, user's accuracy of selection, and more. We experimented with GrabCut, and DEXTR - however, there are just areas where it doesn't really work out for some of our users —
The Need for AI-Guided Labelling
The idea for IntelliBrush came when a few of our enterprise users were painstakingly labelling defective material scans. It took on average 3 minutes to annotate a single stain and 20 minutes to complete an entire image. The pain was obvious. At this point, we knew that we had to implement a faster way to get this done on our annotator.
Wait, but what about outsourcing these tasks to Mechanical Turk or Scale?
Tasks such as identifying tumour lesions in MRI scans to labelling hundreds of different crops are tasks that take years of experience to identify. More often than not, the industry experts are the ones capable of supervising this arduous, painful annotation process.
Will you trust an external labelling workforce to identify tumours and lesions in images like this? Outsourcing doesn't always work, especially in cases like this. The solution is to enable qualified users and professionals to label quickly and accurately at breakneck speeds.
That's why we built IntelliBrush, the fastest in the industry, to help these experts and teams build ground truth 10X faster. Internally.
How Does IntelliBrush Work?
IntelliBrush is built-in to the same web-based annotator on our Nexus platform. It operates on a simple principle - select what is inside and what is outside - and the model will predict the outline of your object. This interaction yields results in less than a second and allows users to make adjustments using left-clicks and right-clicks.
At its core, Intellibrush is a mixture of a few deep neural networks trained on predicting object foreground based on user's clicks. Intellibrush is industry agnostic and works right out of the box for users without any form of pre-training required. This is especially important when building datasets for classes and imageries that contains never-seen-before data. The best part of it all is that we are constantly tuning our IntelliBrush AI, making it better over time! Here's Intellibrush working for various use cases 👇
You can instantaneously label images with a dense amount of objects of varying classes, IntelliBrush analyzes each click and proposes a region where you can iteratively edit until it fits. As an example, lets
IntelliBrush Case Study: Cracked Pavement Detection
Recently, IntelliBrush was used to speed up the annotation process for a crack detection model that was being trained. Before IntelliBrush, one of our user was making use of LabelImg for their mask annotations and labellers were painstakingly clicking around the cracked parameter, bringing the lead time to develop a proof-of-concept model to roughly 6 - 8 weeks.
With IntelliBrush on the Nexus platform, a POC was generated within 1 week, leaving much more room for ML engineers to focus on improving model performance as well as adding in data for edge cases that was causing model confusion.
Data labelling tends to be the bane of many ML teams and is often over-looked, but the age old adage that 'garbage in is garbage out' holds true for ML projects. IntellIBrush aims to make the process a little less painful for teams so that they can focus on other tasks such as model deployments and feature engineering.
What's Next for IntelliBrush?
While the possibilities are endless, we have a roadmap for pushing IntelliBrush further. We are creating a bounding-box version for folks who labels in this format. It works on the same principle and adjustments are made exactly the same. However, instead of generating masks, IntellIBrush will identify the rectangular bounds that best fits the user's object selection.
Want to get started?
IntelliBrush is now available for early access, if you and your team would like to experiment with IntelliBrush, feel free to reach out and apply for early access.
If you have more questions, feel free to join our Community Slack to post your questions. If you have troubles building your own model, simply use our platform, Nexus, to build one in a couple of hours for free.