Depth Anything V3 TensorRT ROS 2 Node

ros2_depth_anything_v3_tensorrt_short.mp4

A ROS 2 node for Depth Anything V3 depth estimation using TensorRT for real-time inference. This node subscribes to camera image and camera info topics and publishes directly both, a metric depth image and PointCloud2 point cloud.

Overview

• Features
• Dependencies
• Topics
• Parameters
• Usage
• Model Preparation
• Docker Image
• Building
• Performance
• Architecture
• Depth Postprocessing Pipeline
• Troubleshooting
• License
• Citation
• Acknowledgements

Features

• Real-time metric depth estimation using Depth Anything V3 with TensorRT acceleration
• Point cloud generation from metric depth image
• Debug visualization with colormap options
• Configurable precision (FP16/FP32)

Important

This repository is open-sourced and maintained by the Institute for Automotive Engineering (ika) at RWTH Aachen University.
We cover a wide variety of research topics within our Vehicle Intelligence & Automated Driving domain.
If you would like to learn more about how we can support your automated driving or robotics efforts, feel free to reach out to us!
📧 opensource@ika.rwth-aachen.de

Dependencies

• Tested with image nvcr.io/nvidia/tensorrt:25.08-py3

  • Ubuntu 24.04, ROS 2 Jazzy
  • CUDA 13
  • TensorRT 10.9

• Tested with image nvcr.io/nvidia/tensorrt:25.03-py3

  • Ubuntu 24.04, ROS 2 Jazzy
  • CUDA 12.8.1
  • TensorRT 10.9

Depending on your driver and CUDA version you need to select the appropriate base image.

Topics

Subscribed Topics

• ~/input/image (sensor_msgs/msg/Image): Input camera image (supports raw and compressed transport via image_transport)
• ~/input/camera_info (sensor_msgs/msg/CameraInfo): Camera calibration info

Published Topics

• ~/output/depth_image (sensor_msgs/msg/Image): Depth image (32FC1 format)
• ~/output/point_cloud (sensor_msgs/msg/PointCloud2): Generated point cloud
• ~/output/depth_image_debug (sensor_msgs/msg/Image): Debug depth visualization (if enabled)

Parameters

All parameters can be configured via config/depth_anything_v3.param.yaml or passed at launch time.

Model Configuration

Parameter	Type	Default	Description
onnx_path	string	"models/DA3METRIC-LARGE.onnx"	Path to Depth Anything V3 ONNX or TensorRT engine file
precision	string	"fp16"	Inference precision ("fp16" or "fp32")

Sky Handling

Parameter	Type	Default	Description
sky_threshold	double	0.3	Threshold for sky classification from model's sky output (lower = more sky detected)
sky_depth_cap	double	200.0	Maximum depth value (meters) to assign to sky pixels

Point Cloud Configuration

Parameter	Type	Default	Description
point_cloud_downsample_factor	int	2	Publish every Nth point (1 = full resolution, 10 = every 10th point)
colorize_point_cloud	bool	true	Add RGB colors from input image to point cloud

Debug Configuration

Parameter	Type	Default	Description
enable_debug	bool	false	Enable debug visualization output
debug_colormap	string	"JET"	Colormap for depth visualization (see below)
debug_filepath	string	"/tmp/depth_anything_v3_debug/"	Directory to save debug images
write_colormap	bool	false	Save colorized debug images to disk
debug_colormap_min_depth	double	0.0	Minimum depth value for colormap normalization (meters)
debug_colormap_max_depth	double	50.0	Maximum depth value for colormap normalization (meters)

Available Colormaps

JET, HOT, COOL, SPRING, SUMMER, AUTUMN, WINTER, BONE, GRAY, HSV, PARULA, PLASMA, INFERNO, VIRIDIS, MAGMA, CIVIDIS

Usage

Basic Launch

ros2 launch depth_anything_v3 depth_anything_v3.launch.py

With Custom Topics

ros2 launch depth_anything_v3 depth_anything_v3.launch.py \
    input_image_topic:=/your_camera/image_raw \
    input_camera_info_topic:=/your_camera/camera_info \
    output_depth_topic:=/depth_anything_v3/depth \
    output_point_cloud_topic:=/depth_anything_v3/points

With Debug Enabled

ros2 launch depth_anything_v3 depth_anything_v3.launch.py \
    params_file:=src/depth_anything_v3/config/debug.param.yaml

Model Preparation

1. Obtain the ONNX model (Two Options): A. Download the ONNX file from Huggingface: https://huggingface.co/TillBeemelmanns/Depth-Anything-V3-ONNX B. Generate ONNX following the instruction here
2. Place model file: Put the ONNX/engine file in the models/ directory
3. Update configuration: Modify config/depth_anything_v3.param.yaml with the correct model path

Expected model input format:

• Input shape: [1, 3, 280, 504] (batch, channels, height, width)
• Input type: float32
• Value range: [0, 1] (normalized with ImageNet mean/std)

Expected model outputs:

• depth: [1, 1, 280, 504] - metric depth map (float32)
• sky: [1, 1, 280, 504] - sky classification logits (float32, lower values = sky)

Docker Image

We precompile and provide a Docker image with all dependencies installed and ready to use. You can pull it from Docker Hub:

docker pull tillbeemelmanns/ros2-depth-anything-v3:latest-dev

If you want to run rosbags and visualize the output in rviz2 make sure to install it

apt update && apt install -y \
ros-jazzy-rviz2 \
ros-jazzy-rosbag2 \
ros-jazzy-rosbag2-storage-mcap

We recommend to use the docker image in combination with our other tools for Docker and ROS.

• docker-ros automatically builds minimal container images of ROS applications
• docker-run is a CLI tool for simplified interaction with Docker images

Building

# From your ROS 2 workspace
colcon build --packages-select depth_anything_v3 --cmake-args -DCMAKE_BUILD_TYPE=Release

source install/setup.bash

# Optional: Pre-build TensorRT engines
./generate_engines.sh

Performance

The node is optimized for real-time performance.

Performance on Quadro RTX 6000:

• DA3METRIC-LARGE: 50 FPS

Architecture

Input Image + Camera Info
         ↓
    Preprocessing (CPU/GPU)
         ↓  
    TensorRT Inference (GPU)
         ↓
    Postprocessing (CPU)
         ↓
   Depth Image + Point Cloud

Depth Postprocessing Pipeline

Depth Anything V3 predicts a dense metric depth map. After TensorRT inference, the node performs the following postprocessing steps:

1. Depth Extraction

• The single-channel NCHW tensor is reshaped into a cv::Mat
• Negative depth values are clamped to zero

2. Focal Length Scaling

• The raw depth output is scaled based on the camera's focal length
• Scale factor: focal_pixels / 300.0 where focal_pixels = (fx + fy) / 2
• This aligns the model's internal focal normalization with your actual camera intrinsics

3. Sky Handling

The model outputs a separate sky classification tensor:

• Pixels with sky confidence below sky_threshold are classified as sky
• Sky pixels are filled with a depth value derived from the 99th percentile of valid (non-sky) depths
• The fill value is capped at sky_depth_cap meters to avoid unrealistic far distances

4. Resolution Upscaling

• The depth map is resized from model resolution back to the original camera resolution using cubic interpolation

5. Point Cloud Generation

For each pixel in the depth map:

X = (u - cx) * depth / fx
Y = (v - cy) * depth / fy
Z = depth

Where (cx, cy) is the principal point and (fx, fy) are the focal lengths from CameraInfo.

Optional features:

• Downsampling: Use point_cloud_downsample_factor to reduce point count for visualization or bandwidth
• Colorization: When colorize_point_cloud is enabled, RGB values from the input image are added to each 3D point

Troubleshooting

Common Issues

1. TensorRT engine building fails:

   • Check CUDA/TensorRT compatibility
   • Verify ONNX model format
   • Increase workspace size

2. Point cloud appears incorrect:

   • Verify camera_info calibration
   • Check coordinate frame conventions
   • Validate depth value units

3. Performance issues:

   • Enable FP16 precision
   • Check GPU memory usage

Debug Mode

Enable debug mode to troubleshoot:

enable_debug: true
debug_colormap: "JET"
debug_colormap_min_depth: 0.0
debug_colormap_max_depth: 50.0
write_colormap: true

This will publish colorized depth images and save them to disk for inspection.

License

This project is licensed under the Apache License, Version 2.0 - see the LICENSE file for details.

Citation

If you use this code in your research, please cite the following:

@misc{beemelmanns2024depth,
  author = {Till Beemelmanns},
  title = {ros2-depth-anything-v3-trt: ROS2 TensorRT Node for Monocular Metric Depth estimation and Point Cloud generation with Depth Anything V3},
  year = {2025},
  publisher = {GitHub},
  url = {https://github.com/ika-rwth-aachen/ros2-depth-anything-v3-trt}
}

Acknowledgements

Thanks to the following repositories for inspiration:

• ROS 2 MoGE TRT Node
• Depth-Anything-V3
• Monocular_Depth_Estimation_TRT
• DepthAnything-ROS