High-performance C++/CUDA + TensorRT implementation of UniDepth V2 with camera intrinsics input and uncertainty/confidence outputs.

• C++ API: include/unidepth_v2.h
• CUDA kernels: kernels/
• CLI demo: examples/image_inference.cpp (built as build/image_inference)
• Python tooling (optional): export ONNX + build TensorRT engines

• RGB image: any size (resized internally to 518x518)
• Camera intrinsics: fx, fy, cx, cy

• Depth: metric depth in meters
• Confidence: higher is better
• Uncertainty: higher is worse (computed as 1 / (confidence + 1e-6))
• Point cloud: PLY with H*W vertices

• FP16 engines
• CUDA Graph support (optional)
• GPU preprocessing / postprocessing (CUDA kernels)

• CUDA Toolkit
• TensorRT (set TensorRT_ROOT if not installed in a default location)
• OpenCV development headers
• CMake >= 3.18
• A C++17 compiler

./build.sh

This produces:

• build/image_inference

./build/image_inference \
  --engine engines/unidepth_v2_vitl14_fp16.trt \
  --input /path/to/image.png \
  --intrinsics "525.0,525.0,319.5,239.5" \
  --output-depth depth.png

Additional optional flags:

• --output-uncertainty uncertainty.png
• --output-confidence confidence.png
• --output-points points.ply
• --no-cuda-graph
• --benchmark N

Example console output (ViT-L14 FP16 engine; values depend on the image and GPU):

========================================
UniDepth V2 TensorRT Inference
========================================
Engine:  engines/unidepth_v2_vitl14_fp16.trt
Input:   assets/demo/rgb.png

Image size: 640x480
Using provided intrinsics:
  fx: 525
  fy: 525
  cx: 319.5
  cy: 239.5

Running inference...
Inference time: 25 ms

Output statistics:
  Depth range:       [1.595, 6.536] meters
  Confidence range:  [0.484, 4.360]
  Uncertainty range: [0.332, 2.064]

Saving outputs...
  Depth visualization: depth.png
  Uncertainty map: uncertainty.png
  Confidence map: confidence.png
Saved point cloud: points.ply (307200 points)

✓ Done!

Generated files (when using default flags):

• depth.png
• uncertainty.png
• confidence.png
• points.ply

This repo intentionally does not publish estimated speeds.

Accuracy summary (PyTorch vs TensorRT FP16, same input):

• Confidence: higher is better
• Uncertainty: higher is worse (1 / (confidence + 1e-6))

Intrinsics are provided as:

• fx,fy,cx,cy (pixel units)

If you don’t pass --intrinsics, the demo estimates intrinsics from image size using a 60° horizontal field-of-view.

python3 python/export_onnx.py \
  --model lpiccinelli/unidepth-v2-vitl14 \
  --output models/unidepth_v2_vitl14_camera.onnx \
  --height 518 \
  --width 518

python3 python/build_engine.py \
  --onnx models/unidepth_v2_vitl14_camera.onnx \
  --output engines/unidepth_v2_vitl14_fp16.trt \
  --precision fp16 \
  --workspace 6144