A client in the medical technology industry has a requirement to protect the anonymity of individuals whose images or videos are being used. This can be achieved by localizing and obfuscating sensitive information present in these media. By obscuring this information, the client aims to preserve the privacy and confidentiality of the individuals, thereby ensuring that their personal information is not disclosed to unauthorized parties. This approach not only helps in maintaining the individuals’ anonymity but also safeguards their rights and dignity.

Main open source libraries: Gluoncv, Opencv, Decord
• Programming language: Python
• Operating system / GPU usage: Ubuntu 20.04.3 LTS / GeForce RTX 2070 Super with Max-Q Design

The use of recorded videos from hospital operating rooms for training machine learning models necessitates the application of full body anonymization to protect the privacy of patients. BMW-Anonymization-Api is a free open-source solution for this task, but its implementation requires significant computing resources and can slow down the anonymization process.

An anonymization API can be built to achieve the same goal of full body anonymization of recorded videos from hospital operating rooms for training machine learning models, while also speeding up the process. This can be accomplished by using a combination of several pretrained models for people segmentation, mask replacement, and video processing tools. Gluoncv provides a wealth of resources for implementing cutting-edge deep learning algorithms in computer vision and the pretrained models can be utilized directly or fine-tuned to meet user specifications. To optimize computing and reduce running time, the API should be executed on a Linux system with GPU support. The use of video processing tools such as OpenCV and Decord can also help increase the performance of the anonymizer.

Some pretrained models from Gluoncv and people segmentation:

+ deeplab_resnet101_voc

+ deeplab_resnet152_voc

+ deeplab_resnet101_coco

+ psp_resnet101_voc

+ psp_resnet101_coco

+ Unet Video processing tools:

+ opencv

+ decord

A high-performing anonymization API has been developed that can anonymize 2 frames of video per second. The models used can be further improved by training them with customer-created datasets, giving the customer the flexibility to control and enhance the performance of the trained models. This customization ability provides the customer with the ability to optimize the API for their specific requirements, resulting in improved results and increased overall satisfaction with the API.