// the find
ika-rwth-aachen/ros2-depth-anything-v3-trt
ROS2 TensorRT Node for Monocular Metric Depth estimation and Point Cloud generation from camera images with Depth Anything V3
A ROS 2 node that runs Depth Anything V3's metric depth model through TensorRT and publishes a 32FC1 depth image and a PointCloud2 from a monocular camera and its CameraInfo. It is aimed at robotics and automated-driving teams that want metric depth from a live camera without writing their own TensorRT pipeline.
- Pre- and postprocessing have dedicated CUDA sources (preprocess_gpu.cu, postprocess_gpu.cu) alongside CUDA error-check and unique_ptr wrappers, so the GPU path is designed in from the start rather than bolted on.
- The depth pipeline is documented stage by stage with the actual formulas: focal scaling, sky fill, cubic upscaling, and back-projection X = (u - cx) * depth / fx. Someone chasing a bad point cloud can check each stage against the doc.
- FP16 and FP32 are both supported, a generate_engines script prebuilds engines, and a Docker image is published. That removes most of the TensorRT setup work for ROS users.
- The focal scaling uses a fixed divisor of 300.0 with no derivation in the README. The metric claim depends on that constant matching the model's training focal length, and nothing here shows it holds across different cameras.
- There are no accuracy numbers. The only performance figure is 50 FPS for DA3METRIC-LARGE on one GPU, with no latency breakdown and no error measurement against ground-truth depth.
- The model input is fixed at 280x504. The README does not say whether camera frames are letterboxed or stretched to that shape. A stretched frame back-projected with the original intrinsics would give a skewed point cloud.
- The tree has no tests directory, and the troubleshooting section is three short bullets. For a node meant to run unattended on a vehicle or robot, that is thin.