Implicit Neural Representations for Video Compression
keywords COMPRESSION, DEEP LEARNING, VIDEO ANALYSIS
Reference persons ENRICO MAGLI
Research Groups CCNE - COMMUNICATIONS AND COMPUTER NETWORKS ENGINEERING, ICT4SS - ICT FOR SMART SOCIETIES, Image Processing Lab (IPL)
Thesis type RESEARCH
Description A new research line in computer vision replaces traditional discrete representations of signals (e.g. pixel grids in images and video) with continuous functions parameterized by deep neural networks. These architectures, called implicit neural representations, take as input the spatio-temporal coordinates and are trained to output a representation of the signal at each input location. Recent studies showed that these representations are a powerful tool, allowing accurate representations of natural signals and offering many possible benefits over conventional representations. In particular, implicit neural representations have been used to represent video, showing promising results. The purpose of this thesis is to investigate the potential of such signal representations in the context of video compression. The student will develop a video encoder based on implicit neural representations, assesing the rate-distortion performance of the proposed encoder with respect to state-of-the-art techniques.
Sitzmann, Vincent, et al. "Implicit Neural Representations with Periodic Activation Functions." arXiv preprint arXiv:2006.09661 (2020). - https://vsitzmann.github.io/siren/
Tancik, Matthew, et al. "Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains." arXiv preprint arXiv:2006.10739 (2020).
Required skills Candidate students should have some background on neural networks. Some experience of TensorFlow environment and Python programming are desirable, along with good programming skills.
Deadline 17/07/2021 PROPONI LA TUA CANDIDATURA