Deploy Depth Anything V3 on Jetson AGX Orin

Depth Anything V3 generates high-quality depth maps from single RGB images. This skill deploys it on Jetson AGX Orin with ROS2 and TensorRT acceleration for real-time robotics use.

Execution model

Run one phase at a time. After each phase:

• Relay all output to the user.
• If output contains [STOP] → stop immediately, consult the failure decision tree.
• If output ends with [OK] → tell the user "Phase N complete" and proceed to the next phase.

Prerequisites

Phase 1 — Install system dependencies (~3 min)

sudo apt update
sudo apt install -y \
    build-essential cmake git libopencv-dev \
    python3-pip python3-colcon-common-extensions v4l-utils
pip3 install numpy opencv-python

[OK] when all packages install without error.

Phase 2 — Configure CUDA environment (~1 min)

echo '
# CUDA Environment
export CUDA_HOME=/usr/local/cuda
export PATH=$CUDA_HOME/bin:$PATH
export LD_LIBRARY_PATH=$CUDA_HOME/lib64:$LD_LIBRARY_PATH
export CUDACXX=$CUDA_HOME/bin/nvcc
' >> ~/.bashrc
source ~/.bashrc
nvcc --version

[OK] when nvcc --version prints CUDA version info. [STOP] if nvcc is not found.

Phase 3 — Install ROS2 camera packages (~2 min)

sudo apt install -y ros-humble-usb-cam ros-humble-cv-bridge \
    ros-humble-image-transport ros-humble-image-geometry ros-humble-rviz2

[OK] when all ROS2 packages install successfully.

Phase 4 — Clone and build the ROS2 workspace (~5 min)

git clone https://github.com/zibochen6/ros2-depth-anything-v3-trt.git
cd ros2-depth-anything-v3-trt
colcon build --packages-select depth_anything_v3 --cmake-args -DCMAKE_BUILD_TYPE=Release
source install/setup.bash

[OK] when colcon build completes with no errors. [STOP] if build fails.

Phase 5 — Download ONNX model and generate TensorRT engines (~10–20 min)

1. Download the .onnx model from https://huggingface.co/TillBeemelmanns/Depth-Anything-V3-ONNX
2. Place it in the ros2-depth-anything-v3-trt/onnx/ directory
3. Generate engines:

chmod +x generate_engines.sh
./generate_engines.sh onnx

[OK] when .engine files appear in the onnx/ directory. [STOP] if conversion fails.

Phase 6 — Camera calibration (optional but recommended)

Install calibration tools:

sudo apt install -y ros-humble-camera-calibration ros-${ROS_DISTRO}-v4l2-camera

Launch camera and calibrate:

# Terminal 1: launch camera
ros2 run v4l2_camera v4l2_camera_node --ros-args -p image_size:=[640,480] -p pixel_format:=YUYV

# Terminal 2: run calibration
ros2 run camera_calibration cameracalibrator \
  --size 8x6 --square 0.025 \
  --ros-args --remap image:=/image_raw --remap camera:=/v4l2_camera

Save calibration parameters to camera_info_example.yaml.

[OK] when calibration completes and parameters are saved.

Phase 7 — Run depth estimation (~1 min)

For USB camera:

CAMERA_INFO_FILE=camera_info_example.yaml ENABLE_UNDISTORTION=1 ./run_camera_depth.sh

For video file:

./run_video_depth.sh

[OK] when depth visualization appears.

Failure decision tree

Reference files

• references/source.body.md — Full Seeed Wiki tutorial with screenshots, calibration details, and video demo links (reference only)