PhD position at Orange: AI for Video Coding with adaptation and competition
1 Avril 2022
Catégorie : Doctorant
A PhD position is proposed at Orange Innovation on Video Coding using AI.
The thesis can start from October 1st, the office is located in Orange Atalante, Rennes, France
Year after year, the amount of video content exchanged on the Internet continues to increase. In addition, new video formats emerge regularly, consistently offering more immersion to the user. As a result, compression tools need to evolve such to adapt to the variety of formats and provide the tools in order to convey videos on the networks. They must offer an ever-lower bit rate, while guaranteeing good quality to the end-user. In addition, it is also needed to compress content of different nature (textures, depth and geometry maps) and to take advantage of the redundancies between these elements.
In 2020, a standard called “Versatile Video Coding” MPEG-I/H.266 was released. It consists of an incremental improvement of previous standards (HEVC/H.265 and AVC/H.264) and significantly increases compression efficiency, both for “standard” 2D video and immersive contents.
Along with the improvement of these conventional compression algorithms, algorithms based on neural networks have emerged in recent years. The approach is to replace the entire encoder-decoder chain with a neural network, which is optimized to minimize the video bit rate while maximizing the quality obtained. In an image compression framework, these end-to-end neural approaches are already able to offer performances comparable to the best conventional compression algorithms. Work is underway to transpose this performance into a video compression framework, and it is envisaged that a standard will be developed from 2025 to reflect the progress in compression. It is in this context that the neural approaches studied in this thesis will play an important role.
One of the weakest points of existing neural approaches is the systematic processing of input data, which does not take into account the varying nature of the signal to be compressed. This is an important difference with conventional approaches (HEVC, VVC) which offer a competition of different coding modes, which are different ways for compressing the signal and adapting to its nature.
The objective of this thesis is thus to massively introduce the notion of competition and content adaptation in a neural compression approach. In this way, the video encoder will be able to offer several coding alternatives to a decoder, making the compression suitable for the content and the target bit rate. As a result, it should be possible to significantly improve the compression performance of neural approaches.
Skills and qualities required
Research Master or engineering school
· Signal processing skills
· Appetite for image / video processing
· In-depth knowledge of Python, C++, Bash etc.
· Experience in machine learning, including deep neural networks (DNN), PyTorch/Tensorflow framework
· Notions of intellectual property (patents)
· Rigor and creativity
Please apply at directly at: