Authors: Christoph Reich, Oliver Hahn, Daniel Cremers, Stefan Roth, Biplob Debnath

Published on: April 18, 2024

Impact Score: 7.6

Arxiv code: Arxiv:2404.12330

Summary

• What is new: The examination of JPEG and H.264 compression on a wide range of vision tasks beyond classification, including localization and dense prediction, highlighting significant accuracy losses.
• Why this is important: The degradation in inference accuracy of deep vision models on edge devices when employing standardized codecs like JPEG and H.264 for data transmission.
• What the research proposes: The paper investigates the implications of using standardized codecs in deep vision pipelines, especially how it affects various deep vision tasks.
• Results: Compression through JPEG and H.264 codecs decreases semantic segmentation accuracy by over 80% in mIoU, indicating severe degradation across multiple vision tasks.

Technical Details

Technological frameworks used: nan

Models used: Deep vision models

Data used: Image and video data using JPEG and H.264 compression

Potential Impact

Companies in the tech industry relying on edge devices or mobile technology for deep vision tasks, particularly those deploying image and video analysis for applications like surveillance, automotive, and mobile applications, could face significant impacts.

Want to implement this idea in a business?

We have generated a startup concept here: OptiVisionAI.