ros2-fast-stereo : ROS2 FastFoundationStereo
A ROS2 workspace for real-time, GPU-accelerated stereo depth estimation from wide-field-of-view or pinhole camera pairs, combining a shared camera-projection abstraction with a TensorRT port of FastFoundationStereo.
This ROS2 (Jazzy) workspace is built from three packages: camera_models, which abstracts seven camera projection models (pinhole, Double Sphere, fisheye polynomial, Kannala-Brandt 8, equirectangular, Mei and Omni) behind one interface, loadable from ROS2 parameters or YAML; perception_utils, which precomputes CPU or CUDA/NPP pixel-to-pixel reprojection maps to rectify arbitrary camera pairs into a shared virtual pinhole view; and ffs_depth, which provides the ROS2 nodes that run the rectification and stereo inference.
Unlike conventional real-time stereo pipelines, which assume a narrow-baseline, forward-facing camera pair with nearly parallel optical axes, this workspace supports wide field-of-view, non-parallel camera pairs as long as their fields of view overlap, deriving the rectifying rotation automatically from each camera's calibrated pose.
Once a synchronized frame pair arrives over a ROS subscription, GPU remap, normalization, TensorRT stereo-matching inference and disparity-to-depth conversion all run on-device through CUDA, NPP and TensorRT, publishing disparity images and colored point clouds.