LET-NET implements an extremely lightweight network for feature point extraction and image consistency computation. The network can process a 240 x 320 image on a CPU in about 5ms. Combined with LK optical flow, it breaks the assumption of brightness consistency and performs well on dynamic lighting as well as blurred images.

1. The LET-NET training code is released at https://github.com/linyicheng1/LET-NET-Train.
2. Gray Image is also suport in LET-NET, you can get pytorch and onnx model tpye in ./model/
3. LET-VINS Demo run on UMA-VI dataset is released.
4. Our proposed LET-VINS won the second place in the VIO track of the ICCV2023SLAM Challenge, which is the best performance among the traditional methods.
5. The preprinted paper was posted at here.

• Breaking of brightness consistency in optical flow with a lightweight CNN network,Yicheng Lin, Shuo Wang, Yunlong Jiang, Bin Han, arXiv:2310.15655, pdf

• OpenCV (https://docs.opencv.org/3.4/d7/d9f/tutorial_linux_install.html)
• ncnn (https://github.com/Tencent/ncnn/wiki/how-to-build#build-for-linux)

Notes: After installing ncnn, you need to change the path in CMakeLists.txt

set(ncnn_DIR "<your_path>/install/lib/cmake/ncnn" CACHE PATH "Directory that contains ncnnConfig.cmake")

mkdir build && cd build
cmake .. && make -j4

You can enter the path to a video or two images.

./build/demo <path_param> <path_bin> <path_video>

or

./build/demo <path_param> <path_bin> <path_img_1> <path_img_2>

For example using the data we provide:

./build/demo ./model/model.param ./model/model.bin ./assets/nyu_snippet.mp4

You should see the following output from the NYU sequence snippet:

The left is ours and the right is the original optical flow algorithm.

The left is ours and the right is the original optical flow algorithm.

The left is ours and the right is the original optical flow algorithm.

@ARTICLE{let-net,
  author={Lin, Yicheng and Wang, Shuo and Jiang, Yunlong and Han, Bin},
  journal={IEEE Robotics and Automation Letters}, 
  title={{Breaking of brightness consistency in optical flow with a lightweight CNN network}}, 
  year={2024},
  pages={1-8}
}