Deep Learning for Make & Model Recognition – Why?
Our brain is much advanced in visual recognition. And therefore we probably don’t know why Deep Learning for Make & Model recognition is even needed.
Well, of course a human brain can effortlessly distinguish between
- a panda and a polar bear,
- can read a number plate or
- recognize the face of another human
But these tasks are fairly difficult for a computer. It only seems easy, because our brain is excellent in understanding images.
In the last couple of years the Machine or Deep Learning has begun to takle those difficulties.
And especially due to a type of model called Convolutional Neural Networks (CNN), now reasonable resolutions for difficult visual recognition tasks are receivable.
The ablilities of these networks are not only close to those of humans, but even excel in certain areas. One example from the image processing field would be the recognition of car maker and type, also called Make & Model recognition (MMR).
„Make“ in the phrase „Make & Model“ refers to the maker of the vehicle such as e.g.: Ford, Toyota, Honda etc. „Model“ refers to the specific name the maker as given the car type such as e.g.: Focus, Accord, Pathfinder, etc.
For human brains this seems a straight forward task, especially for car aficionados. Main characteristics such as, logos, hood ornaments, or labels make it easy for humans to identify the car. But for a computer this was always a difficult task, due to the visual complexity of a car or other vehicles.
What is the power of Make & Model?
It can not “only” recognize care types and distinguish between Audi and BMW. It can also recognize the year of construction of the same model. A difficult task for the human brain.
Of course not only cars of the same maker look simillar but also the models of SUVs of differen makers or different Sedans of Audi or BMW. Make & Model recognition, therefore not only sees the obvious. Cars such as a Bugatti Veyron “Hypercar” are of course easy to distinguish. But also differences which are not as obvious are recognized with Deep Learning. For example different rear wings (fastbacks) with makers such as Nissan and Toyota, which look quite simillar. They can even be distinguished. And that even on images captured on a street with cars passing by, on highways and other highly frequented cars.
We would now say: “But there is always a logo on the car”. True, but what if we cannot see the logo? And what about the year of construction and model of cars popular in Europe, Asia or US? Can we recognize those as well?
There are even limits for car aficionados.