ROS2 Monodepth

Note: This project was generated with the assistance of artificial intelligence.

This ROS2 project implements monocular depth estimation using the ZoeDepth model. It provides real-time depth estimation from single RGB images.

Installation

1. Install dependencies:

rosdep install -i --from-path src --rosdistro humble -y --ignore-src

2. Build the workspace:

cd ~/ros2_ws
colcon build --symlink-install
source install/setup.bash

Usage

1. Start the webcam publisher node:

ros2 run zoedepth webcam_publisher --ros-args -p device_id:=0

2. In another terminal, start the depth estimator node:

ros2 run zoedepth depth_estimator --ros-args -p compiler_backend:='aot_eager'

Configuration

Both nodes support various parameters that can be set via the command line:

Webcam Publisher Parameters:

• device_id (default: 0): Webcam device ID or path
• target_width (default: 256): Target width for resizing
• target_height (default: 256): Target height for resizing
• force_square_crop (default: false): Force square output by cropping to shortest dimension before resizing
• publish_rate (default: 15.0): Publishing rate in Hz

Depth Estimator Parameters:

• model_repo (default: 'isl-org/ZoeDepth'): Model repository
• model_type (default: 'NK'): Model type (N, K, or NK)
• normalize_depth (default: false): Whether to normalize depth output to 0-255 range
• colorize_output (default: false): Whether to apply colorization to the depth map using magma colormap
• measure_latency (default: false): Whether to measure and log processing latency
• use_compiler (default: true): Whether to use PyTorch's compiler
• compiler_backend (default: 'inductor'): Compiler backend to use. Options:
  • 'inductor': Default PyTorch 2.0 compiler
  • 'eager': Traditional PyTorch eager execution
  • 'aot_eager': Ahead-of-time compilation with eager execution
  • 'tensorrt': TensorRT acceleration (requires torch-tensorrt package)

Note: To use the TensorRT backend, you must first install the torch-tensorrt package:

pip3 install torch-tensorrt

Performance

Performance measurements were conducted on an NVIDIA RTX 4070 GPU. Initial testing shows:

• Average latency of ~120ms per frame at 8Hz with no compiler backend
• Similar performance with Inductor or TensorRT backend in initial tests, so no visible acceleration
• Testing on NVIDIA Jetson Orin platforms is planned

These numbers are preliminary and may vary based on your specific hardware configuration and input resolution.

Topics

• /image_raw (sensor_msgs/Image): Raw RGB images from webcam
• /depth/image_raw (sensor_msgs/Image): Estimated depth maps

License

This project is licensed under the MIT License - see the LICENSE file for details.