Democratization of Computer Vision


Cheap computers, cameras and lenses

Back in 2015, we wanted to load 8 cameras in a high altitude balloon. We used some Raspberry Pi model A+ and some official Raspberry Pi cameras. The stock cameras had a field of view that was a bit too narrow for what we wanted. So we used magnetic, wide angle lenses designed for smartphones. This gave us some options.


Most people discard conversion lens as junk. And optically, they are terrible, with lots of barrel distortion and other issues. In fact, barrel distortion is easy to see on videos on YouTube taken with wide (it is the built in lens actually) lens "extreme sports" type of video cameras that are so popular. The closer to the edge, the more it distorts.

But this was not going to be a show stopper for us. With the right software, I was convinced we could get some really decent shots. So I mounted a camera with a wide angle lens on top of the raspberry pi case, powered the pi with my phone spare battery pack (with a micro USB lead) and took pictures to calibrate my software (written in Python) and then went downtown Winston Salem to take pictures of buildings with lots of horizontal and vertical lines. The conversion software worked extremely well.

Even more choices

Today, the Raspberry Pi officially launched a new camera module. The original Raspberry Pi Camera V.1 was released in 2013 and had a resolution of 5 megapixels (in both regular and infrared versions). The V.2 was released in 2016 with a resolution of 8 megapixels (again, in regular and infrared versions). The new module is the HQ Camera, with 12.3 megapixels. It also has a threaded mount, compatible with camera tripods. But beyond the increase in resolution, it has a feature that is really valuable in research: interchangeable lenses.



Check out the getting started guide, the brief and the mechanical drawing pdfs.

Lens Mount

The mount for the interchangeable lenses is a C mount (technically, it is a CS mount and with a spacer, it is compatible with a C mount). C mount lenses were introduced by Bell and Howell for their 16mm film cameras, in the 1920s, as a replacement to their B mount. Compatible lenses are still in manufacture in both C and the newer CS mount and are readily available in zoom and prime lenses.

Some fairly fast 25, 35 and 50mm C mount (for 1" sensors)
Some CCTV 2/3 lenses in 25, 35 & 50mm C mount with macro spacer, shield & cap

In a future blog post, I'll provide some examples of performance and interaction in a Jupyter Notebook.

A guide with practical examples

Up to now, with just hardware, we couldn't really justify our blog post title of "Democratization of Computer Vision". But the Raspberry Pi foundation didn't release just a camera module, or a video4linux driver. They launched a guidebook: "The Official Raspberry Pi Camera Guide for Camera Module & High Quality Camera". It is available as a printed book that can be purchased, or as a free PDF download.

The first few chapters will get you acquainted with the camera and utilities like raspistill and raspivid, and controlling the camera from Python. There are also plenty of projects to keep anyone busy for a while. But quite interestingly enough, there are also 2 chapters that are really bringing Computer Vision to the masses. For example, Chapter 11 has the heading:

Chapter 11
Car Spy Pi
Who's that parked on the driveway?
Find out automatically using ANPR

ANPR is automatic number plate recognition. If students get into that type of Computer Vision projects today, imagine what they'll be doing in the future! (Just don't go and sell this as a secure solution to Sheffield...)

More on Computer Vision

A few weeks back, a friend contacted me to get some information on some specific aspect of Computer Vision. In the process, I raided my bookshelves and came up with a list of various resources (classes, books, papers etc) which I'll share in a future post. Of course, one or two blog post will only scratch the surface. We are not going to state that we can cover it all as a summer project [1]...

Francois Dion

Chief Data Scientist

@f_dion


About Dion Research LLC: We are a boutique Data Science consultancy . established 2011. As we do end-to-end Data Science, we can help you solve business problems every step of the way. Get in touch for more information.

[1] MIT. The Summer Vision Project, Seymour Papert, 1966 (pdf)

Comments

Post a Comment