libmaskrcnn-trt

A C++17 library for easily running Mask R-CNN using NVIDIA TensorRT.

Build

Requirements:

• An NVIDIA GPU (tested on an NVIDIA Jetson Xavier NX).
• A CUDA installation (tested on CUDA 10.2).
• A TensorRT installation (tested on TensorRT 7.1.3).
• A C++17 compatible compiler (tested on gcc 7.5.0 on Ubuntu 18.04).
• OpenCV 4
• CMake
• Make (optional, for convenience)

To install CUDA and TensorRT follow the instructions from NVIDIA. To install the other dependencies on Debian/Ubuntu run

sudo apt install g++ make cmake libopencv-dev

Then run make to build the library and the example executables.

Usage

To use the library you'll need to first download the Mask R-CNN UFF model mrcnn_nchw.uff by running make download-model (requires wget). This model was generated using these instructions.

Using the library in your code

The recommended way is to incorporate this repository into your project (e.g. as a git submodule) and then add it as a subdirectory in your CMakeLists.txt, e.g.:

add_subdirectory(libmaskrcnn-trt)

# Build a CMake target with libmaskrcnn-trt.
target_include_directories(PUBLIC some-target libmaskrcnn-trt/include)
target_link_libraries(some-target maskrcnn-trt)

# Make sure to change the path to the libmaskrcnn-trt directory in
# add_subdirectory() and target_include_directories() if needed.

Example executable

An example executable is compiled along with the library. To run inference on two images you can run

./build/release/maskrcnn-trt-example mrcnn_nchw.uff /PATH/TO/IMAGE2 /PATH/TO/IMAGE2

Information about the detected objects will be shown on standard output and visualizations will be created in /PATH/TO/IMAGE1.detections.png and /PATH/TO/IMAGE2.detections.png. The images can be in any format supported by OpenCV's cv::imread().

Notes

• The first time the Uff model is loaded it will be converted into a device-specific format. This process can take a few minutes. To speed-up subsequent runs it is possible to serialize the device-specific format to the disk by setting mr::MaskRCNNConfig::serialized_model_filename to the name of a non-existent file. The serialized version will be loaded on subsequent runs instead of doing the conversion each time.
• The first inference can take up to 2x more time than subsequent inferences.
• On newer versions of TensorRT some of the functions used in libmaskrcnn-trt have been deprecated. The code was retained as is for compatibility with TensorRT 7 which is the only version currently officially supported on the Jetson Xavier NX. To suppress the deprecation warnings you can pass the -Wno-deprecated-declarations option to the compiler.

License

Copyright 2019 NVIDIA Corporation
Copyright 2021 Smart Robotics Lab, Imperial College London
Copyright 2021 Sotiris Papatheodorou, Imperial College London

Distributed under the Apache License Version 2.0

The library uses code from this sample in the TensorRT repository.

Many thanks to ivanhc for their help in this issue.