Umedia: För den som gillar medieteknik

 

Projects

"Digital djurpark"

digital djurpark








Very low bitrate video coding based on principal component analysis

Why use arbitrary frequencies to represent a video when you know what the video contains? We use images of a persons face to build a model of this person's facial mimic. With the model it is possible to encode video into extremely low bitrates (< 5 kbps) and reconstruct this video with very high quality. It is even possible to encode video with high definiton (HD) resolution into these low bitrates. To make the model efficient we make use of hands-free equipment. This equipment enables a camera to follow the movement of the person using the equipment; the user can roam freely and use both hands. Wearable video equipment

 

The equipment in the figure above is from Easyrig AB. Photo: Kalle Prorok



A model consists of several images of the same person with different facial expressions. According to the American psychologist Paul Ekman it is possible to model all facial expressions with just six basic emotions. The basic expressions are happiness, fear, sadness, surprise, disgust and anger.

The basic expressions

Happiness, fear, sadness, surprise, disgust and anger.



Encoding and decoding of video with this model is very simple compared to standard video encoding. Encoding and decoding is simply multiplications, subtractions and additions. The low complexity saves energy, which is a very important factor for battery-driven devices. The very low bitrate also saves energy since the device uses less power for transmission.



WAWO.NET: A facial similarity search engine

WAWO, which stands for "Digg Me" in Chinese, is a fully automated content based face image retrieval system (search engine). The system has been under development in the Digital Media Lab in Umeå University over the past five years and has now reached the general availability release stage (see publication [1.1]). It relies on human facial characteristics and it uses a database of non-annotated images, where some humans may or may not appear. Given the above, the proposed system can search for a particular person when the initial query is another image (content based searching).

query image

In the above figure we see a user supplied image query along with the detected face whereas in the following image one can see the resulted images:

results

For some old projects you can go here