Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Exploring AI-based Anonymization of Industrial Image and Video Data in the Context of Feature Preservation

Sabrina Cynthia Triess, Timo Leitritz, Christian Jauch

TL;DR

This paper evaluates DeepPrivacy2 for full-body anonymization of industrial image and video data, aiming to protect privacy while preserving utility for downstream tasks. It compares DP2 against conventional anonymization (blur/pixelation) across pose estimation and action recognition, using PTK, EPE, PTA, and ACC metrics. The results show DP2 generally preserves more pose information and action-detection capability than blur or pixelation, with near-100% PTK and $EPE$ under 17 pixels in several scenarios, but facial expressions, hand poses, and temporal consistency can be affected. The findings inform practical privacy decisions in industrial settings and suggest DP2 as a potential data-augmentation tool, while noting limitations for precise hand- or facial-gesture tasks and the need for improved temporal stability.

Abstract

With rising technologies, the protection of privacy-sensitive information is becoming increasingly important. In industry and production facilities, image or video recordings are beneficial for documentation, tracing production errors or coordinating workflows. Individuals in images or videos need to be anonymized. However, the anonymized data should be reusable for further applications. In this work, we apply the Deep Learning-based full-body anonymization framework DeepPrivacy2, which generates artificial identities, to industrial image and video data. We compare its performance with conventional anonymization techniques. Therefore, we consider the quality of identity generation, temporal consistency, and the applicability of pose estimation and action recognition.

Exploring AI-based Anonymization of Industrial Image and Video Data in the Context of Feature Preservation

TL;DR

This paper evaluates DeepPrivacy2 for full-body anonymization of industrial image and video data, aiming to protect privacy while preserving utility for downstream tasks. It compares DP2 against conventional anonymization (blur/pixelation) across pose estimation and action recognition, using PTK, EPE, PTA, and ACC metrics. The results show DP2 generally preserves more pose information and action-detection capability than blur or pixelation, with near-100% PTK and under 17 pixels in several scenarios, but facial expressions, hand poses, and temporal consistency can be affected. The findings inform practical privacy decisions in industrial settings and suggest DP2 as a potential data-augmentation tool, while noting limitations for precise hand- or facial-gesture tasks and the need for improved temporal stability.

Abstract

With rising technologies, the protection of privacy-sensitive information is becoming increasingly important. In industry and production facilities, image or video recordings are beneficial for documentation, tracing production errors or coordinating workflows. Individuals in images or videos need to be anonymized. However, the anonymized data should be reusable for further applications. In this work, we apply the Deep Learning-based full-body anonymization framework DeepPrivacy2, which generates artificial identities, to industrial image and video data. We compare its performance with conventional anonymization techniques. Therefore, we consider the quality of identity generation, temporal consistency, and the applicability of pose estimation and action recognition.
Paper Structure (11 sections, 2 equations, 5 figures)

This paper contains 11 sections, 2 equations, 5 figures.

Table of Contents

  1. INTRODUCTION
  2. RELATED WORK
  3. EVALUATION METHODS
  4. Anonymization
  5. Pose Estimation
  6. Action Recognition
  7. EXPERIMENTAL RESULTS
  8. Visual Quality of DeepPrivacy2
  9. Pose Estimation on Anonymized Images
  10. Action Recognition on Anonymized Videos
  11. CONCLUSION

Figures (5)

  • Figure 1: Pose estimation on original, DeepPrivacy2-anonymized, blurred and pixelated images (from left to right). The "Office" scenario is shown.
  • Figure 2: End-Point Error per scenario and anonymization type at a score threshold of 0.5 with respect to PTK. Each marker represents one image with at least one detected keypoint.
  • Figure 3: Action distribution of 21,111 video frames in total at a score threshold of 0.2. Only the top five actions are shown.
  • Figure 4: Percentage of Total Actions per anonymization type of 21,111 video frames in total.
  • Figure 5: Action accuracy per anonymization type of 21,111 video frames in total as a function of the score threshold.