NVIDIA Project Redtail team published code, models and samples for depth estimation from stereo using NVStereo DNN. Check out our GitHub: https://github.com/NVIDIA-Jetson/redtail/tree/master/stereoDNN
In short, this work contains:
- Pre-trained TensorFlow and corresponding TensorRT models that provide dense depth estimation from stereo camera
- Inference library that implements some of the TensorFlow functionality (e.g. 3D [de]convolutions) as TensorRT C++ API plugins
- Sample application that demonstrates how to use the code
- Scripts to convert NVStereo DNN TensorFlow models to TensorRT C++ API
TensorFlow is not required to run the inference (see sample app for more details).
This video https://youtu.be/0FPQdVOYoAU provides an overview of the work.
Feel free to provide feedback/ask questions on GitHub issues page.