Depth Anything 3 ROS2 Wrapper

Camera-agnostic ROS2 wrapper for Depth Anything 3 monocular depth estimation.

Performance (2026-02-05)

Platform	Backend	Model	Resolution	FPS
Orin AGX 64GB	PyTorch FP32	DA3-Small	518x518	~5
Jetson Orin NX 16GB*	TensorRT FP16	DA3-Small	518x518	23+ / 43+

23+ FPS real-world (camera-limited), 43+ FPS processing capacity

*Tested on Seeed reComputer J4012

Jetson Users: Host requires numpy, pycuda, and TensorRT Python bindings. ./run.sh auto-installs these.

---

Key Features

• TensorRT-Optimized: 40+ FPS on Jetson via TensorRT 10.3
• Camera-Agnostic: Works with any camera publishing ROS2 image topics
• One-Click Demo: ./run.sh handles everything automatically
• Shared Memory IPC: Low-latency host-container communication (~8ms)
• Multiple Models: DA3-Small, Base, Large with auto hardware detection
• Docker Support: Pre-configured for Jetson deployment

---

Quick Start

Option 1: Jetson TensorRT Demo (Recommended)

git clone https://github.com/GerdsenAI/Depth-Anything-3-ROS2-Wrapper.git ~/depth_anything_3_ros2
cd ~/depth_anything_3_ros2
./run.sh

First run takes ~15-20 minutes (Docker build + TensorRT engine). Subsequent runs start in ~10 seconds.

Options:

./run.sh --camera /dev/video0   # Specify camera
./run.sh --no-display           # Headless mode (SSH)
./run.sh --rebuild              # Force rebuild Docker

Option 2: Native ROS2 Installation

# Clone and install
git clone https://github.com/GerdsenAI/GerdsenAI-Depth-Anything-3-ROS2-Wrapper.git
cd GerdsenAI-Depth-Anything-3-ROS2-Wrapper
bash scripts/install_dependencies.sh
source install/setup.bash

# Run with USB camera
ros2 launch depth_anything_3_ros2 depth_anything_3.launch.py \
  image_topic:=/camera/image_raw

See Installation Guide for detailed steps.

Option 3: Docker (Desktop GPU)

docker-compose up -d depth-anything-3-gpu
docker exec -it da3_ros2_gpu bash
ros2 launch depth_anything_3_ros2 depth_anything_3.launch.py image_topic:=/camera/image_raw

See Docker Guide for more options.

---

Architecture

This project uses a host-container split for optimal Jetson performance:

HOST (JetPack 6.x)
+--------------------------------------------------+
|  TRT Inference Service (trt_inference_shm.py)    |
|  - TensorRT 10.3, ~15ms inference                |
+--------------------------------------------------+
                    ^
                    | /dev/shm/da3 (shared memory)
                    v
+--------------------------------------------------+
|  Docker Container (ROS2 Humble)                  |
|  - Camera drivers, depth publisher               |
|  - SharedMemoryInferenceFast (~8ms IPC)          |
+--------------------------------------------------+

Why: Container TensorRT bindings are broken in current Jetson images. Host TensorRT 10.3 works perfectly.

---

Platform Recommendations

Platform	Model	Resolution	Expected FPS	Memory
Orin Nano 4GB/8GB	DA3-Small	308x308	40-50	~1.2GB
Orin NX 8GB	DA3-Small	308x308	50-55	~1.2GB
Jetson Orin NX 16GB*	DA3-Small	518x518	43+ (validated)	~1.8GB
AGX Orin 32GB/64GB	DA3-Base	518x518	25-35	~2.5GB

*Validated on Seeed reComputer J4012

See Optimization Guide for detailed benchmarks and tuning.

---

Topics

Subscribed

Topic	Type	Description
~/image_raw	sensor_msgs/Image	Input RGB image

Published

Topic	Type	Description
~/depth	sensor_msgs/Image	Depth map (32FC1)
~/depth_colored	sensor_msgs/Image	Colorized visualization (BGR8)
~/confidence	sensor_msgs/Image	Confidence map (32FC1)

---

Common Parameters

Parameter	Default	Description
model_name	depth-anything/DA3-BASE	Model to use
device	cuda	cuda or cpu
inference_height	518	Input resolution height
inference_width	518	Input resolution width
publish_colored	true	Publish colorized depth

See Configuration Reference for all parameters.

---

Documentation

Guide	Description
Installation	Detailed installation steps, offline setup
Usage Examples	USB camera, ZED, RealSense, multi-camera
Configuration	All parameters, topics, models
ROS2 Node Reference	Node lifecycle, QoS, Jetson performance tuning
Optimization	Platform benchmarks, performance tuning
Jetson Deployment	TensorRT setup, host-container split
Docker	Container deployment options
Troubleshooting	Common issues and solutions

---

Requirements

• ROS2: Humble Hawksbill (Ubuntu 22.04)
• Python: 3.10+
• TensorRT: 10.3+ (Jetson JetPack 6.x) for production
• CUDA: 12.x (optional for desktop GPU)

---

Acknowledgments

• Depth Anything 3 - ByteDance Seed Team (paper)
• NVIDIA TensorRT - High-performance inference
• Jetson Containers - dusty-nv's L4T Docker images
• Hugging Face - Model hosting

Inspired by grupo-avispa/depth_anything_v2_ros2 and scepter914/DepthAnything-ROS.

---

Citation

@article{depthanything3,
  title={Depth Anything 3: A New Foundation for Metric and Relative Depth Estimation},
  author={Yang, Lihe and Kang, Bingyi and Huang, Zilong and Zhao, Zhen and Xu, Xiaogang and Feng, Jiashi and Zhao, Hengshuang},
  journal={arXiv preprint arXiv:2511.10647},
  year={2024}
}

---

License

This ROS2 wrapper: MIT License

Depth Anything 3 models:

• DA3-Small: Apache-2.0 (commercial use OK)
• DA3-Base/Large/Giant: CC-BY-NC-4.0 (non-commercial only)

---

Contributing

Contributions welcome! We especially need help with test coverage for the SharedMemory/TensorRT production code paths. See CONTRIBUTING.md for:

• Current test coverage status
• Priority areas needing tests
• How to write and run tests

---

Support

• GitHub Issues
• GitHub Discussions
• Troubleshooting Guide