skills.rest
NVIDIA
NVIDIAOFFICIAL
Explore all skills by this author →
WebsiteGitHub

tao-train-foundation-stereo

Train, evaluate, and deploy FoundationStereo models for stereo disparity estimation with NVIDIA TAO.

3.2k|370|Updated Feb 25, 2026
Search Tags:#model training#tensorRT#3D reconstruction#nvidia tao#stereo depth#disparity estimation#foundationstereo
One-click install
npx skills add https://github.com/NVIDIA/skills --skill tao-train-foundation-stereo
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: tao-train-foundation-stereo
Source: https://github.com/NVIDIA/skills/tree/main/skills/tao-train-foundation-stereo
Command: npx skills add https://github.com/NVIDIA/skills --skill tao-train-foundation-stereo

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

Training and deploying stereo depth estimation models requires navigating complex TAO Toolkit configuration specs, dataset annotation formats, and containerized GPU workflows. This Skill guides an AI agent through the complete FoundationStereo lifecycle so disparity maps can be produced from stereo image pairs without manual trial and error.

Core Features & Use Cases

  • Full Model Lifecycle: Covers train, evaluate, inference, export, and quantize actions through the unified TAO depth_net CLI, plus TensorRT deployment via TAO Deploy.
  • Dataset Guidance: Maps data layouts (Middlebury, KITTI, ETH3D, FSD, IsaacReal, Crestereo) to the correct dataset class and annotation file format, with a convert step to generate annotations.
  • AutoML Routing: AutoML-enabled train action routes through tao-run-automl by default, with per-run automl_policy overrides for plain training.
  • Use Case: A robotics engineer with KITTI stereo pairs asks the agent to train a FoundationStereo model; the agent generates the annotation file, writes the spec YAML with the Kitti dataset class, launches the Docker training run, and verifies KPIs in status.json.

Quick Start

Ask your agent to train a FoundationStereo stereo depth model on your stereo image dataset using NVIDIA TAO.

Frequently Asked Questions about tao-train-foundation-stereo

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I train a FoundationStereo model with NVIDIA TAO?▼

Create an annotation file listing left, right, and disparity image paths, write a spec YAML from the skill's override template with model_type FoundationStereo and the matching dataset_name, then run depth_net train inside the TAO Docker container with your data and output directories mounted.

What dataset formats does TAO FoundationStereo support?▼

Supported dataset classes include Middlebury, Kitti, Eth3d, FSD, IsaacRealDataset, Crestereo, and a GenericDataset fallback. Annotation files use 2 columns for inference, 3 with ground-truth disparity, or 4 with an occlusion mask.

Can I export FoundationStereo to TensorRT for deployment?▼

Yes, but not through the PyT depth_net entrypoint, which rejects gen_trt_engine. Export to ONNX with the export action, then follow the TAO Deploy workflow in references/tao-deploy-foundation-stereo.md to build TensorRT engines and run TensorRT evaluate or inference.

Why does my FoundationStereo smoke test fail with shape mismatches?▼

Shape mismatches occur when crop_size is smaller than max_disparity or uses 112x112 crops. The skill validates smoke runs at crop_size 128x128 with both dataset.max_disparity and model.max_disparity set to 128.

Does FoundationStereo training support AutoML hyperparameter optimization?▼

Yes, the model is AutoML-enabled and train requests route through tao-run-automl by default when the train schema and spec template are packaged. Set automl_policy to off for a single run to use direct training instead.

What are the hardware requirements for running TAO depth_net?▼

You need Docker with nvidia-container-toolkit and an NVIDIA GPU. The container is launched with --gpus, 16G shared memory, and writable home and cache directories mounted under the output path when running as a non-root user.

Related Skills

Explore More Software Engineering Skills→
wangyendtwangyendt
8

pywayne-cv-stereo-tag-matcher

Match AprilTags across stereo camera image pairs for correspondence visualization.

Community
Intermediate
infantesromeroadrianinfantesromeroadrian

inference-optimization

Quantize production inference models and export to ONNX with TensorRT optimization.

Community
Advanced
NVIDIA-TAONVIDIA-TAO
83

tao-launch-workflow

Automate TAO workflow launches with platform selection, credentials, and preflight checks.

Official
Advanced
NVIDIA-TAONVIDIA-TAO
83

tao-train-depth-anything-v2

Train monocular depth models using DepthAnything v2 architecture.

Official
Advanced