DepthAnything-ROS

DepthAnything-ROS is 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 environment

• Install ROS2

See ROS2 document.

To install ROS2 easily, I recommend to use ansible script of Autoware. In detail, please see the installation page.

• Install dependency

sudo apt install libgflags-dev libboost-all-dev

• Prepare your rosbag

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

Set onnx files

Set onnx files for DepthAnything-ROS/data or set onnx_path parameter.

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

Run below command to get the onnx files of pre-train model.

# Install gdown
pip install gdown
# Download onnx file
mkdir data && cd data
gdown 1jFTCJv0uJovPAww9PHCYAoek-KfeajK_

If you want to make onnx files at yourself, please use depth-anything-tensorrt.

Launch

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 made by DepthAnything.

Parameters

• onnx_path (string)
  • Default parameter: "$(find-pkg-share depth_anything)/data/depth_anything_vitb14.onnx"

The path to onnx file.

• precision (string)
  • Default parameter: "fp32"

The precision mode to use quantization. DepthAnything-ROS supports in "fp32" or "fp16" (#2) for now.

Note

Build for TensorRT engine

When you run on the first start up, you need to wait about 5 minutes for build step.

Performance

• Performance
  • RTX4090 results is written in official code

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

Reference

• Depth-Anything
• trt-depth-anything Use for TensorRT inference
• depth-anything-tensorrt Use for making from pth files to onnx files
• tensorrt_common Use library
• cuda_utils Use library
• gdown