ROS2 Wrapper for Depth Anything 3
Monocular depth estimation for robotics applications with real-time performance and GPU acceleration.
overview
depth anything 3 ros2 is a wrapper that brings ai-powered monocular depth estimation to ros2 robotics environments. it enables real-time depth perception from rgb camera streams, making it ideal for autonomous navigation, obstacle avoidance, and 3d scene understanding in robotic applications.
the wrapper integrates seamlessly with existing ros2 camera pipelines and supports multiple model variants optimized for different hardware configurations, from edge devices to high-performance gpu workstations.
features
- real-time depth estimation: process rgb images to generate accurate depth maps in real-time
- multiple model variants: choose from small, base, large, or giant parameter models
- gpu acceleration: cuda support for high-performance inference on nvidia gpus
- configurable parameters: customize topics, device selection, and model variants at runtime
- ros2 integration: standard sensor_msgs/image compatibility with existing pipelines
performance
model performance characteristics on nvidia rtx 3090:
| model | parameters | vram | fps |
|---|---|---|---|
| small | 80m | ~2gb | ~45 |
| base | 120m | ~3gb | ~35 |
| large | 350m | ~5gb | ~25 |
| giant | 1.15b | ~12gb | ~10 |
system requirements
- ros2: humble or newer
- python: 3.10 or higher
- pytorch: 2.0 or higher
- cuda: optional, for gpu acceleration
installation
1. install pytorch and python dependencies:
pip install torch torchvision opencv-python2. install depth anything 3:
pip install git+https://github.com/DepthAnything/Depth-Anything-V2.git3. build the ros2 package:
cd ~/ros2_ws/src
git clone https://github.com/spectateursimon/Depth-Anything-3-ROS2.git
cd ~/ros2_ws
colcon build --packages-select depth_anything_3_ros2usage
basic launch with default parameters:
ros2 launch depth_anything_3_ros2 default.launch.pylaunch with custom parameters:
ros2 launch depth_anything_3_ros2 default.launch.py model_name:=depth-anything/DA3-Small device:=cpuparameters
| parameter | default | description |
|---|---|---|
| image_topic | /camera/color/image_raw | input rgb camera stream |
| depth_image_topic | /depth | output depth image topic |
| device | cuda:0 | computation device (cuda:0, cpu) |
| model_name | depth-anything/DA3-Large | model variant to use |
use cases
- autonomous navigation: enable depth perception for mobile robots and drones
- obstacle avoidance: detect and avoid obstacles in real-time using monocular cameras
- 3d scene understanding: generate depth maps for slam and mapping applications
- research and development: prototype and test depth estimation algorithms in ros2
credits
this ros2 wrapper is based on depth anything 3 by bytedance. depth anything 3 is a state-of-the-art monocular depth estimation model that achieves exceptional performance across diverse scenarios.