LiKL-Cpp

Cpp implementation of "LiKL: A Lightweight Approach for Joint Detection and Description of Keypoint and Line". Which is capable of simultaneously extracting keypoint and line features in an image and inferring feature descriptors in a single process.

The pytorch implementation in LiKL

Updates

2024-04-28

• Add TensorRT support. (Only support static input shapes now.)

Usage

1. Dependency

• Eigen3

sudo apt install libeigen3-dev

• google-glog

sudo apt install libgoogle-glog-dev

• OpenCV4
• libtorch

wget -O libtorch.zip https://download.pytorch.org/libtorch/cu113/libtorch-cxx11-abi-shared-with-deps-1.11.0%2Bcu113.zip
unzip libtorch

• torchvision

wget https://github.com/pytorch/vision/archive/refs/tags/v0.12.0.zip
unzip vision-0.12.0.zip

## Build torchvision
cd vision
mkdir build && cd build
cmake .. -DWITH_CUDA=on -DUSE_PYTHON=on -DCMAKE_PREFIX_PATH=/path/to/libtorch
make
make install

• TensorRT (Optional) Tested in 8.6.1.6

2. Build

mkdir build && cd build
# Add -DWITH_TENSORRT=on support for the TensorRT if needed
cmake .. -DCMAKE_PREFIX_PATH=/path/to/libtorch
make

3. Download the converted model

TorchScript model

You can download from Baidu (y3yu) or Onedrive.

You can also convert pytorch model to torchcript. Please refer to LiKL.

ONNX model

You can convert pytorch model to onnx. Please refer to LiKL.

4. demo

Set the torchscript path in configs/likl.yaml

./build/test_likl ./asset/terrace0.JPG ./asset/terrace1.JPG

5. Acknowledgements

Thanks to SuperPoint-SuperGlue-TensorRT)