WIT is specialized in image processing software and can assist in building or creating applications in the field of industrial automation, visual inspection, 3D reconstructions from video images, image compression, medical image analysis, product inspection, barcode reading, ANPR applications, etcetera.
WIT has initiated several projects and joined many others in a variety of application areas.
Some examples of this are briefly highlighted here.
Sphericam
Recently, Wit joined
Sphericam.inc
. as a co-owner, to develop the core software for its main innovation, the
Sphericam
.
Six cameras in one faceted ball fuse six videos in one single "all around"
video.
The
wonderful apparatus
is the brainchild of the inspirational Jeffrey Martin, a 360 and ultra high resolution panorama photographer who is intrigued by 360 degree video. Servaes Joordens designs the camera hardware for it.
We hope this product will boost production of 360 video's, which can be viewed by VR systems such as the
Oculus Rift
and others but also with your smartphone. The
Sphericam
may provide innovative ways of shooting movies, because it makes choosing the camera direction obsolete.
This
i-have-recorded-all
feature of the Spericam makes for endless possibilities in film editing.
And because the ball has all the positional sensors on board, the movie always knows its orientation, speed and position. This facilitates easy production of horizontally stabilized videos or in the direction you move, for example.
Imagine, with this cam on your helmet you can ski downhill and just enjoy the ride without worrying about your video at all. You can produce a forward looking video later, perfectly stabilized with your movements, with the horizon or even compass direction if you wish!
Or you can just re-view the entire 360 degrees with your VR headset or mobile phone, and discover why that friend behind you suddenly hit the snow.
Pipeline welding
In cooperation with Flatlands B.V. a wireless, portable camera was developed to view and record steel oil pipe welding. The camera makes use of an
NIT sensor
that cannot overexpose. It uses a constant shutter time but is able to record the entire range of illuminations at once, in a 16 bit WDR image. The software processes the raw images to give the best view per single image in the video stream, so there is no blinding effect when the welding starts or black screen when the welding stops.
Road inspection
Earlier, WIT produced software for
Vision Components GmbH
, based in Karlsruhe Germany. This software is used in road inspection cameras for the US police for vehicle detection and reading of license plates. It can run on any ARM core processor.
New projects
Some time ago,
Yann Lecun
did important pioneering work on so-called deep-learning. His
convolutional neural networks
are able to recognize the content of images. But d
eep networks are very large and take lots of time to train.
It would be nice it it was faster and less complicated!
Questions are
* Does stripping the convolution training out of it still work?
* Can these layers be replaced by predefined filters?
* Can relative position indicators replace the
max-pooling
layers for translation invariance?
If these question can be answered positively, pixels might still know about object edges and their own position.
If indeed the case, the potential could be enormous. Deep network performance in an easy jacket would open up practical applications directly. Think of symbol detection, face recognition, etcetera.
Research on this subject is one of WIT's aims.
As a practical test case, WIT wants to create a pixel-based object classifier, using an "upside-down" approach by first doing classification and then grouping into objects.
In license plate reading for example, this software could easily "see through" dirt, shadows and other disturbances. In the computer simulated example image below, this is visualized.
Another example of that pixel-based classifier, could be vehicle tracking. This might be done by simply "learning" vehicle features like the front window edge, as in the simulated example video below.