branding

TSA 3D Image Segmentation Application

Role

Dillon - Product Architect

Timeline
Sep - Dec 2021
abstract project cover

The Client.

The client was the Transportation Security Administration (TSA).

The Problem.

Need to accelerate processing times at TSA checkpoints. We were tasked with creating a machine learning model to detect and segment prohibited items (i.e. guns, knives, etc…) in a 3D baggage scan.

The Solution.

No dataset existed to train the model on. So guess what, we made one. Does this mean we went out and spent tens of thousands of $s to have data collected? Absolutely not. We did it the minimalist way, by creating a synthetic training dataset to fine-tune a pre-trained model on.


You see, ML models don't know the difference between real world data and synthetic data (most of the time). With that, we took 3D models of prohibited items and injected them into real baggage scenes with proper material values (given by MRI machine). This, combined with data augmentation techniques, allowed us to get a working model going within the span of 2 weeks.

The Process.

I start by asking, what can we do with a minimal amount of resources. Owing to my start-up experience, I know that a LOT of work that is done is often unnecessary and wasteful. I cut out everything that's non-essential and start there.

The Result.

We had a working model in 2 weeks. Way before they expected, and even more importantly, way before they needed.

Ready to work with us?

Let's get to work

Ready to work with us?

Let's get to work

Ready to work with us?

Let's get to work

© 2022 Standard Data, Inc.
© 2022 Standard Data, Inc.
© 2022 Standard Data, Inc.