Source: Deep Learning on Medium
How NOUS uses SSIM to make Annotating Easy
NOUS is an effective tool because it allows users to annotate images easily, and therefore, train algorithms with minimal effort. For more complex datasets where there are many small annotations to make, training can still be tedious for the expert. Annotating repetitive objects, such as beehive structures, small bacterial colonies, or animals in a drone image can take hours of focused work.
In an extra effort to make NOUS the most efficient deep learning tool, SSIM has been integrated into the app in the form of a “Detection Assistant” to aid with annotations. This will save experts hours of annotating items individually, and allow a single input to result in many annotations.
What is SSIM?
Structural Similarity Index is a perceptual metric that compares a processed image with its original, and measures the difference. In NOUS, the user provides a reference patch by drawing a bounding box around the object of interest. This is then compared with the rest of the image, building up a heat-map of high and low similarity regions which are used to select possible candidates for areas containing objects similar to that in the reference patch.
How does it work?
The SSIM “Detection Assistant” processes a single annotation and finds visually similar structures in the image. This reduces annotation time significantly. A slider is used to increase the number of detections to encompass an optimal number of correct annotations. The assistant can then be applied again to find additional structures with slightly different characteristics.
What does it mean for users?
By integrating SSIM into the NOUS platform, annotation time is minimized for even the most detailed datasets. The aim of NOUS is to put powerful technology into the hands of the expert so they can train the algorithm to do the repetitive work, and allow the expert more time to focus on tasks that require their specialized skills. With SSIM, users are enabled to reduce annotation time and begin to tackle problems that, in the past, were simply too large.
View the full tutorial below.