TensorRT Depth Anything for Efficient Inference

city_depth_anything_1024x256.mp4

Purpose

This package estimates depth for arbitrary images using TensorRT Depth Anything for efficient and faster inference. Specifically, this supports multi-precision and multi-device inference for efficient inference on embedded platforms.

Setup

Install basic libraries for inference on GPUs including CUDA, cuDNN, TensorRT (>=8.6), and OpenCV.

Also, you'll need to install them as below.

sudo apt-get install libgflags-dev
sudo apt-get install libboost-all-dev

ONNX Conversion

Please use the export_to_onnx.py from the following repository for the converter.

https://github.com/spacewalk01/depth-anything-tensorrt

Build sources

git clone [email protected]:tier4/trt-depth-anything.git
cd trt-depth-anything.git
cd build/
cmake ..
make -j

Start inference

-Build TensoRT engine

./trt-depth-anything --onnx depth_anything_vitb14.onnx --precision fp32

-Infer from a Video

./trt-depth-anything --onnx depth_anything_vitb14.onnx --precision fp32 --v {VIDEO PATH} --depth {gray/magma/jet} (--max_depth VALUE)

-Infer from images in a directory

./trt-depth-anything --onnx depth_anything_vitb14.onnx --precision fp32 --d {Directory PATH} --depth {gray/magma/jet} (--max_depth VALUE) (--save_detections --save_detections_path {SAVE_PATH}) (--dont_show)

Cite

Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, Hengshuang Zhao, "Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data", arXiv:2401.10891, 2024 [ref]

Parameters

--precision={fp32,fp16}

--v={VIDEO PATH}

--d={Directroy PATH}

--depth={gray/magma/jet} : default is 'gray'

--max_depth={VALUE} : option

Assumptions / Known limits

Todo

• Confirm accuracy using INT8
• Support Multi-batch execution

Reference repositories

• https://github.com/LiheYoung/Depth-Anything
• https://github.com/spacewalk01/depth-anything-tensorrt
• https://github.com/autowarefoundation/autoware.universe/tree/main/perception/tensorrt_yolox
• https://github.com/autowarefoundation/autoware.universe/tree/main/common/tensorrt_common
• https://github.com/autowarefoundation/autoware.universe/tree/main/common/cuda_utils