While digital content has become widespread in many parts of life, privacy-related concerns have become more significant. This work addresses privacy preservation challenges in apartment walkthrough videos, which are used for marketing and other commercial purposes, using deep learning and computer vision methods.
The thesis broadly covers video and image censoring of specific objects by employing two deep learning semiconnected strategies: object detection and subsequent filtering and object inpainting, which allows one to preserve the integrity of the background. In addition to evaluating the aforementioned methods, critical attention is paid to the computational performance and temporal consistency of the type of video input.
The methodology involves selecting pretrained deep learning models for different strategies from publicly available datasets, as well as a set of metrics used to compare their performance. The effectiveness of these models is evaluated using a specifically defined set of criteria assessing censoring accuracy, including false positive/negative rates, temporal stability, and inference speed.
The results of this study demonstrate the potential of deep learning approaches to enhance privacy in video and imagery content, providing a solid introduction to further developments in this area. Although central attention is paid to the censoring objective, this work also examines application aspects of censoring on modern hardware in production, including framework-specific optimization.
