On Annotation-free Optimization of Video Coding for Machines

Windsheimer M, Brand F, Kaup A (2024)

Publication Type: Conference contribution, Conference Contribution

Publication year: 2024

Pages Range: 1857-1863

Event location: Abu Dhabi KR

ISBN: 979-8-3503-4939-9

URI: https://ieeexplore.ieee.org/document/10647318

DOI: 10.1109/ICIP51287.2024.10647318

Open Access Link: https://arxiv.org/abs/2406.07938

Abstract

Today, image and video data is not only viewed by humans, but also automatically analyzed by computer vision algorithms. However, current coding standards are optimized for human perception. Emerging from this, research on video coding for machines tries to develop coding methods designed for machines as information sink. Since many of these algorithms are based on neural networks, most proposals for video coding for machines build upon neural compression. So far, optimizing the compression by applying the task loss of the analysis network, for which ground truth data is needed, is achieving the best coding performance. But ground truth data is difficult to obtain and thus an optimization without ground truth is preferred. In this paper, we present an annotation-free optimization strategy for video coding for machines. We measure the distortion by calculating the task loss of the analysis network. Therefore, the predictions on the compressed image are compared with the predictions on the original image, instead of the ground truth data. Our results show that this strategy can even outperform training with ground truth data with rate savings of up to 7.5 %. By using the non-annotated training data, the rate gains can be further increased up to 8.2 %.

Authors with CRIS profile

Marc Windsheimer Lehrstuhl für Multimediakommunikation und Signalverarbeitung (LMS) Fabian Brand Lehrstuhl für Multimediakommunikation und Signalverarbeitung (LMS) Andre Kaup Lehrstuhl für Multimediakommunikation und Signalverarbeitung (LMS)

How to cite

APA:

Windsheimer, M., Brand, F., & Kaup, A. (2024). On Annotation-free Optimization of Video Coding for Machines. In Proceedings of the International Conference on Image Processing (ICIP) (pp. 1857-1863). Abu Dhabi, KR.

MLA:

Windsheimer, Marc, Fabian Brand, and André Kaup. "On Annotation-free Optimization of Video Coding for Machines." Proceedings of the International Conference on Image Processing (ICIP), Abu Dhabi 2024. 1857-1863.

BibTeX: Download