Dear all,
I have developed a small Python package to integrate the NVIDIA Fast Foundation Stereo model into ROS2. This model was released in March 2026.
My package subscribes to ROS2 rectified stereo images and camera info, and then uses the NVIDIA Fast Foundation Stereo model to compute a disparity map. My package then publishes a disparity map and a point cloud.
In terms of functionality it is a bit like the Semi Global Block Matching algorithm in stereo_image_proc in ROS2. However this is a neural network approach and performs better (in my opinion) especially on low texture regions of an image.
It requires an NVIDIA card (not necessarily very recent, mine is from 2018). My robot does not have an NVIDIA card, so I demonstrate with my robot publishing images across Wifi to a desktop computer which does have an NVIDIA card (RTX 2070).
YouTube link: ROS2 Fast Foundation Stereo
Github link: ROS2 Fast Foundation Stereo
I hope this may be of interest to the community.