Modelling models

Object detection is a fundamental technology for image analysis, and a key enabler for (semi-)autonomous systems. Object detection systems have advanced rapidly over the past decade, with common objects now being trivial to detect with very little code or experience required.

However, for more bespoke use cases the challenge still exists. Gathering data of niche objects sets and making models that work in resource constrained environments (such as edge processing aboard a UAV) are still active research challenges. This is an area we commonly encounter in defence applications.

So how can we gather data of niche objects? One approach is to go into the field and capture data. This is costly, time consuming, and difficult to organise depending on the object. But what if we had a model of the object?

Enter my low budget photography studio.

My lightbox, turntable, and model tank setup achieved for under £30.

Using a scale model in a lightbox with a turntable, we can easily capture multi perspective video of objects of interest. With a little magic, we can then build a 3D model from this video using a Neural Radiance Field (NeRF). These allow us to create novel views of existing objects

We can do the same in thermal infrared. Obviously here our thermal signature won’t be anything like the real thing (I just warmed the model in the sun for this), but as a proof of concept we’re able to generate a NeRF from it at least.

What are we going to do with these, and can we use any other types of sensors? We’ll discuss more soon!

Previous
Previous

NWRDSC Membership

Next
Next

A quanta of success