Streetview Segmentation

Perform semantic segmentation on images, using Facebook's Mask2Former models.

29 commits | Last commit 17 months ago

What Streetview Segmentation can do for you

Streetview Segmentation script performs semantic segmentation on images, using models from Facebook's collection of Mask2Former models. Output consists of a CSV-file detailing, for each input image, the number of pixels in each semantic class in that image. Optionally, it can save a copy of each input image overlaid with the semantic segmentation. If the input images are 360° photo's, the script provides the possibility of tranforming them by projecting them onto a cube, resulting in six images per input image.

This script is specifically designed to run on a computer without a GPU. Some of the underlying libraries require the presence of CUDA-drivers to run, even if the actual device is absent. As it can be problematic to install such drivers on a computer without an actual GPU, the program is packaged as Docker-container, based on an official NVIDIA-image, which comes with pre-installed drivers. Building and running the container requires the presence of Docker engine. Note that the container is based on a Linux image (Ubuntu 18.04); running it on a Windows computer may require extra configuration.

Programming languages
  • Python 95%
  • Dockerfile 5%
Not specified
</>Source code

Participating organisations

Utrecht University


Maarten Schermer