FoundationStereo
huggingface.co- Category
- Developer Tools
- Rank
- No. 2179Tools index
Previous survey · No. 2070 ·
- Pricing
- Free
- Type
- TOOL
- Use case
- Models: Train & Run
- Date
About
FoundationStereo is an NVIDIA Research foundation model for stereo depth estimation that takes an RGB stereo image pair and outputs an accurate disparity map with strong zero-shot generalization across scenarios. It combines a frozen pretrained DepthAnythingV2 monocular depth network with a side-tuning EdgeNeXt-based CNN, targeting 3D perception in robotics, industrial, and smart-space applications, and is licensed for commercial use under the NVIDIA Open Model License.
What it does
NVIDIA's Hugging Face card presents a network that reads a left and right camera frame and returns per-pixel disparity, which the card says can be turned into depth using camera intrinsics. According to the card, it targets edge deployment on Jetson and other NVIDIA GPUs through TensorRT, ONNX or PyTorch.
Stated on the product site
- License
- NVIDIA Open Model License
- Commercial use
- Stated as ready for commercial use
- Runtimes
- PyTorch, ONNX and TensorRT
- Hardware
- NVIDIA GPUs, including Jetson
- Operating system
- Linux
- Input
- Two RGB stereo images
Not stated on the site
- The card does not show accuracy results on the Middlebury, ETH3D or KITTI evaluation sets it lists.
- Whether the model runs on non-NVIDIA hardware is not addressed on the page.
- The card does not say where training or fine-tuning code for this variant lives.
Written from the product site at huggingface.co.
Tags
Media
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.