DepthAnything-ROS

DepthAnything-ROS is a ROS2 wrapper for Depth-Anything.

nuscenes_demo.mp4

Environment

• Ubuntu 22.04.01, ROS2 Humble
• CUDA 12.3, cuDNN 8.9.5.29-1+cuda12.2, TensorRT 8.6.1.6-1+cuda12.0

Get Started

Set Up the Environment

• 1. Install ROS2

See the ROS2 installation documentation for detailed instructions. To install ROS2 easily, I recommend using the Ansible script from Autoware. For more details, refer to the installation guide.

• 2. Install dependencies

sudo apt install libgflags-dev libboost-all-dev

• 3. Prepare your rosbag

If you don't have a rosbag, I recommend using rosbag for Nuscenes dataset.

• Set ONNX Files

Place the ONNX files in the DepthAnything-ROS/data directory, or set the onnx_path parameter in the launch file.

<arg name="onnx_path" default="$(find-pkg-share depth_anything)/data/depth_anything_vitb14.onnx" />

To download the ONNX files for the pre-trained model, run the following commands:

# Install gdown
pip install gdown

# Download ONNX file
mkdir -p data && cd data
gdown 1jFTCJv0uJovPAww9PHCYAoek-KfeajK_

If you prefer to create the ONNX files yourself, you can use depth-anything-tensorrt.

Launch the node

ros2 launch depth_anything depth_anything.launch.xml

Interface

Input

• input/image (sensor_msgs::msg::Image)
  • The input image.

Output

• ~/output/depth_image (sensor_msgs::msg::Image)
  • The depth image generated by DepthAnything.

Parameters

• onnx_path (string)
  • Default: $(find-pkg-share depth_anything)/data/depth_anything_vitb14.onnx
  • The path to the ONNX file.
• precision (string)
  • Default: "fp32"
  • The precision mode to use for quantization. DepthAnything-ROS supports "fp32" or "fp16" (#2).

Notes

Building the TensorRT Engine

The first time you run the system, you may need to wait for about 5 minutes for the build process to complete.

Performance

• Performance data is based on results from an RTX4090 and RTX2070.

Model	Params	RTX4090 TensorRT	RTX2070 TensorRT
Depth-Anything-Small	24.8M	3 ms	27 ms, VRAM 300MB
Depth-Anything-Base	97.5M	6 ms	65 ms, VRAM 700MB
Depth-Anything-Large	335.3M	12 ms	200 ms, VRAM 1750MB

References

• Depth-Anything
• trt-depth-anything – Use for TensorRT inference
• depth-anything-tensorrt – Use for converting from PTH files to ONNX files
• tensorrt_common – Utility library
• cuda_utils – Utility library
• gdown – Tool for downloading files from Google Drive