Offene Arbeiten

GAN-Based Vibrotactile Signal Compression

Stichworte:
Haptics, Deep Learning
Kurzbeschreibung:
Adaptation of Image Compression Methods to Vibrotactile Signals

Beschreibung

The student will adapt GAN-based compression methods for images such that it can be applied on vibrotactile signals. Vibrotactile signals are single- or multichannel time series data representing vibrations.

Classical compression methods for vibrotactile signals already exist [1], as well as some Deep Learning approaches. Using GANs is a promising next step to increase the perceived quality of compressed signals.

The basis of the work will be HiFiC [2]. A repository by the authors is available at [3] in Tensorflow, but using the repository at [4] in Pytorch would make the adaption with existing code easier.

The main challenges are:

- understanding the existing pipeline for images

- adapting the code to work on single-channel data

- replacing parts of the training loss and compression pipeline to fit the application

- optimization of the pipeline

 

[1] A. Noll, L. Nockenberg, B. Gülecyüz, and E. Steinbach, “VC-PWQ: Vibrotactile Signal Compression based on Perceptual Wavelet Quantization,” in 2021 IEEE World Haptics Conference (WHC), 2021, pp. 427–432. doi: 10.1109/WHC49131.2021.9517217.

[2] F. Mentzer, G. Toderici, M. Tschannen, and E. Agustsson, “High-Fidelity Generative Image Compression.” arXiv, Oct. 23, 2020. Accessed: Feb. 12, 2024. [Online]. Available: http://arxiv.org/abs/2006.09965

[3] “HiFiC.” [Online]. Available: https://github.com/tensorflow/compression/tree/master/models/hific

[4] “HiFiC Pytorch Implementation.” [Online]. Available: https://github.com/Justin-Tan/high-fidelity-generative-compression

Voraussetzungen

Experience in Deep Learning and Python programming

Betreuer:

Lars Nockenberg