SUCCESS STORY

Image Segmentation

Challenge

Our client visually digitizes roads and elements of urban property into 360° images. These 3D scans are used by urbanists, architects, and other system operators to improve the efficiency of their interventions on the field. To increase the usefulness of this data, we have developed a tool that allows for: - The automatic segmentation of the elements found in the images. - The blurring of sensitive data (e.g.: faces). - The computation of the distances between the different urban property elements.

Before we had developed the API for our client, human intervention was necessary to label the images and blur sensitive data. We developed an algorithm that allowed for automatic segmentation and blurring of the images. Furthermore, we also set up a platform that allowed our client’s teams to be fully autonomous in their use of the algorithm.

project weeks

3 weeks

for a launch in production

tailor-made platform

Our Impact

95%

Precision on the blurring of sensitive data

images processed per second

Thanks to Image Recognition, we can now fully automate time-consuming and repetitive tasks, such as segmentation or detection of urban property, so as to better map them.

Geoffroy de Boissieu, Project Director @Sicara

Objectives

Segmentation of panoramic images with heavy GDPR restrictions

Objectives

Segmentation of panoramic images with heavy GDPR restrictions

When roads and public spaces are mapped in 3D, sensitive data is also digitized. To make sure companies remain in compliance with GDPR guidelines, our tool automatically anonymizes the data. From faces to licence plates, our algorithm can blur sensitive personal data with 95% precision and 95% recall.

An optimized customer experience

A platform tailored for business-specific needs

An optimized customer experience

A platform tailored for business-specific needs

To allow for the full autonomy of our client’s teams in their management of the API, we also developed a platform with which they can: - Easily launch algorithm trainings on new batches of datasets. - Label the results returned by the algorithm to correct the potential errors.