Depth-Anything TensorRT in TouchDesigner

TouchDesigner implementation for Depth Anything and Depth Anything v2 with TensorRT monocular depth estimation.

Features

• One click install and run script
• In-TouchDesigner inference

Usage

Tested with TouchDesigner 2023.11880 & Python 3.11

Installation process:

1. Install Python 3.11;
2. Install CUDA Toolkit 11.8;
3. Install GIT;
4. Clone TDDepthAnything repository
5. Download Depth-Anything model or Depth-Anything v2 model
6. Run install.bat. When prompted to select version of Python, type in "3.11" and hit "Enter";
7. When installation will be finished, copy model to checkpoints folder

Acceleration process:

1. Run accelerate.bat.
2. Select model version (1 - Depth-Anything, 2 - Depth-Anything v2)
3. Select model size (s - small, b - base, l - large, g - giant)
4. Select width & height (default is 518x518)
5. Wait for acceleration to finish

TouchDesigner inference:

1. Add TDDepthAnything.tox to project
2. On Settings page change path to TDDepthAnything folder and click Re-init
3. On Depth Anything page select path to engine file (for ex. engines/depth_anything_vits14.engine) and click Load Engine

Acknowledgement

Based on the following projects:

• Depth-Anything - Unleashing the Power of Large-Scale Unlabeled Data.
• Depth-Anything TensorRT C++ - Leveraging the TensorRT API for efficient real-time inference.
• TopArray - Interaction between Python/PyTorch tensor operations and TouchDesigner TOPs.