An AI tool for vision applications in advanced R&D

The new first step for virtually all computer vision problems is a key to unlocking massive gains in advanced R&D.

Your data is basically perfect, right?

The task was a fairly standard computer vision problem for a client.

The objective? Build a computer vision algorithm that can calculate the dimensions of an object in front of a checkerboard with known checkerboard square sizes.

Count a few checkers, make some adjustments for various distortions, and the problem is solved.

Also, mapping out a checkerboard in an image is well-documented within the computer vision community with just a few constraints:

You get the idea. Data quality is usually somewhere in the range from non-ideal to bad. This is especially true in advanced R&D where anomalies are the norm, not the exception, and experimental data point quantity is often low.

... enter SAM from Meta.

SAM as a Computer Vision Tool: Accurate, Tunable, Robust

SAM is an accurate and tunable model that Meta trained on over 1-billion masks on 11-million images. It takes an input of an image (and some hints if you'd like to provide them) and outputs masks of all of the objects of interest in an image.

What are objects of interest? Anything that a human annotator would consider an object in the image.

SAM as applied to a cityscape: from https://segment-anything.com/

SAM has proven to be stable when applied across variable conditions and has demonstrated the ability to mask objects that it never saw when being trained.

If results like those can be generated in a stable manner, why shouldn't it be the starting point for every computer vision analysis problem? Instead of just RGB pixels, you get to start with a set of masks (as a human would define them) within an image. I'll take it.

Build better R&D solutions

If you are spending hours analyzing imagery of any kind with inconsistent results from manual annotation, SAM may be a good solution for your R&D processes.

You can start by reading a more technical guide we put together here:

We'd love to help you build better R&D systems: info@depotanalytics.co