Code Monkey home page Code Monkey logo

iln's Introduction

ILN - Implicit LiDAR Network: LiDAR Super-Resolution via Interpolation Weight Prediction

This project covers the LiDAR super-resolution task, but without the resolution constraint of a deep network. The Implicit LiDAR Network (ILN) reconstructs the dense LiDAR points based on the sparse sensor measurements, depending on the system's requirement. It can predict the detection distance of any continuous query laser within the sensor's field of view. You can see technical details on the project page.

DEPENDENCY

The implementation is mainly based on the PyTorch framework for network training. Also, We use the ROS for 3D visualization and other robotic applications, but it can be optional.

The "requirements.txt" file includes the major Python packages we used. Check the dependencies or install them in your environment using the command:

pip3 install -r python_src/requirements.txt

Our package uses the Pybind11 library to call several C++ voxelization functions from Python code, and is included when cloning with the --recursive option: git clone [email protected]:PinocchioYS/iln.git --recursive

Use these commands to build the voxelization module in a ROS workspace:

cd <your_catkin_repo>
git clone [email protected]:PinocchioYS/iln.git --recursive src/iln
catkin_make

Use these command to build the voxelization module outside or a ROS workspace:

git clone [email protected]:PinocchioYS/iln.git --recursive mkdir build cd build cmake .. make

CARLA DATASET

We collected the training dataset by LiDAR simulation using the CARLA. This dataset consists the 9-realistic outdoor scenes, and each scene has LiDAR data with four different resolutions: 16x1024, 64x1024, 128x2048, and 256x4096 (vertical x horizontal angles). You can download our dataset from the link or the following command.

cd python_src/dataset
wget https://sgvr.kaist.ac.kr/~yskwon/papers/icra22-iln/Carla.zip
unzip Carla.zip

These commands put the dataset into "python_src/dataset/Carla" which our example codes use as the default directory.

PRE-TRAINED MODELS

This project provides several models that we trained with the CARLA dataset: ILN [Ours, ICRA22], LIIF [CVPR21], and LiDAR-SR [RAS20]. You can download the model files from the link or the following commands.

cd python_src/models/trained
wget https://sgvr.kaist.ac.kr/~yskwon/papers/icra22-iln/trained.zip
unzip trained.zip

These commands make the pre-trained models into "python_src/models/trained" where our default model "iln_1d_400.pth" is located.

VISUALIZATION

Visualization is supported with ROS in the implicit_lidar_network package:

usage: demo_resolution_free_lidar.py [-h] -i INPUT_FILENAME -cp CHECKPOINT [-v VOXEL_SIZE]
demo_resolution_free_lidar.py: error: the following arguments are required: -i/--input_filename, -cp/--checkpoint

iln's People

Contributors

pinocchioys avatar juliangaal avatar

Stargazers

Hello World! avatar fanx042 avatar Bin Yang avatar avatar donguk avatar Praveen Kumar Rajendran avatar avatar

Watchers

James Cloos avatar avatar Chen Yao avatar

Forkers

pandaypr gundalpha kimsoohwan

iln's Issues

Number of samples confusion

Hi Youngsun Kwon,

I have a question regarding the training process. I saw from your config here: iln_1d.yaml you set the num_of_samples is equal to 16384. When I check inside the train_models.py and print the shape of the input, output, and pred images all of them has this shape: [b, 16384]. I was thinking why the output has the same shape with the input while the output should have 128x2048 points, right ? Does it mean the model only learn from limited random data from sampling instead of all of the data?

Will the training process be better if I increase the num_of_samples to the number of output pointcloud?

I hope hearing from you soon. Thanks

Another lidar type support

Hi Youngsun Kwon,

I hope you are well.
This week, I was very happy because I found your publication about "Implicit Lidar Network" and I'm very interested to your work after I read your paper thoroughly and your code a little bit. I hope I can read and try your full code soon after you completed this github repo. I want to learn more about your projects. May I ask you several questions about your projects?

In your opinion, does the ILN network work with non-uniform lidar data distribution, such as robosense lidar that has a dense resolution in the middle of the vertical fov? What should I do inside the network to make it works with different kind of lidar data?

Thank you very much. I hope to hear from you soon.

Range Image generator

Hi Youngsun Kwon,

Thanks for the readme, the dataset, and all of the pretrained weights. It is so helpful.

I'm just wondering how did you create the range image from pointclouds? How's your method? If I remember correctly, there are two different method to represent pointcloud to a range image; PBEA (Projection Based Elevation Angle) and PBID (Projection Based Laser ID). Which one do you use? How did you manage to minimize the quantization error.

I tried to convert the range image that I built using my own generator, but It seems it has a bit error as you can see from the image below ( the white points are the raw PCL from rosbag, the green points are the raw range image).

image

Hope hearing from you soon.

Recommend Projects

  • React photo React

    A declarative, efficient, and flexible JavaScript library for building user interfaces.

  • Vue.js photo Vue.js

    🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.

  • Typescript photo Typescript

    TypeScript is a superset of JavaScript that compiles to clean JavaScript output.

  • TensorFlow photo TensorFlow

    An Open Source Machine Learning Framework for Everyone

  • Django photo Django

    The Web framework for perfectionists with deadlines.

  • D3 photo D3

    Bring data to life with SVG, Canvas and HTML. 📊📈🎉

Recommend Topics

  • javascript

    JavaScript (JS) is a lightweight interpreted programming language with first-class functions.

  • web

    Some thing interesting about web. New door for the world.

  • server

    A server is a program made to process requests and deliver data to clients.

  • Machine learning

    Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.

  • Game

    Some thing interesting about game, make everyone happy.

Recommend Org

  • Facebook photo Facebook

    We are working to build community through open source technology. NB: members must have two-factor auth.

  • Microsoft photo Microsoft

    Open source projects and samples from Microsoft.

  • Google photo Google

    Google ❤️ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.