Feature map compression for neural networks
Im Rahmen des Projekts Next Generation Video Coding


Under the topic of Video Coding for Machines (VCM), MPEG aims to standardize a bitstream that is composed of highly compressed, pre-extracted features and an optional video stream. The purpose of this project is the research of highly effective algorithms for the compression of feature streams, which will be part of MPEGs standardization efforts.
Feature streams consist of temporally successive feature map frames. Feature maps are outputted by intermediate or output layers of neural networks. They contain task-specific information. As an example, a possible task would be object detection. In this case, the feature map might contain bounding box coordinates for detected objects.
In this student project, you will have the opportunity to analyze feature streams of typical neural networks regarding their statistical characteristics. Based on these findings you will develop efficient coding schemes, aiming to outperform modern video compression standards such as HEVC (also known as H.265) and VVC (H.266).


Good programming skills in python and (optionally) C++ are mandatory. Knowledge from the machine learning and source coding (Quellencodierung) classes might be helpful.
If you are interested in this topic, feel free to contact me.

Contact person: Martin Benjak