Coder Social home page Coder Social logo

lidarclip's Introduction

LidarCLIP

This is the official implementation of the WACV 2024 paper LidarCLIP or: How I Learned to Talk to Point Clouds.

LidarCLIP overview

Checkpoints

CLIP Version Training Dataset Link
ViT-L/14 ONCE google drive
ViT-B/32 ONCE google drive

Instructions

  • download the SST submodule git submodule update --init --recursive
  • build the dockerfile docker build -t lidarclip -f docker/Dockerfile .
  • spin up a container with access to the dataset and at least one gpu. (docker run -v <lidarclip-path>:/lidarclip <...> lidarclip)
  • in the docker container, change directories to /lidarclip
  • in the docker container, run python train.py --datadir=<dataset-path> --checkpoint-save-dir=<checkpoint-dir> --name=<experiment-name>. You can specify many additional flags, here is an example command: --name lidarclip-main --batch-size 128 --workers 4 --checkpoint-save-dir /proj/lidarclip/checkpoints/ --clip-model ViT-L/14.
  • pre-compute LidarCLIP features: python.py scripts/cache_embeddings.py --checkpoint=<checkpoint-dir>/<checkpoint-file>. By default these are for the once validation set, but it is easily changed with the --dataset-name, --data-path, and --split arguments.
  • now you can explore the capabilities of the model by running one of the notebooks, placed under notebooks/:
    • retrieval_and_zero_shot.ipynb allows for qualitative exploration of retrieval and zero-shot classification
    • retrieval_metric.ipynb is for computing quantitative retrieval metrics
    • diffusion/CLIP_Guided_Stable_Diffusion.ipynb is for lidar-to-image generation
    • generate_captions.ipynb is for lidar captioning. Note that you need to train LidarCLIP against ViT-B/32 (instead of the default which is ViT-L/14).

Note that some paths have to be modified in the notebooks to point your desired dataset and cached features.

Dataset preparation

We refer to the official download and preparation instructions for ONCE and NuScenes. Once the dataset directories are set up according to the official instructions, our code works without any additional steps. Note that NuScenes is entirely optional and only used for evaluating domain shift capabilities.

Citation

If you find this work useful, please cite

@inproceedings{lidarclip2024,
  title={LidarCLIP or: How I Learned to Talk to Point Clouds},
  author={Hess, Georg and Tonderski, Adam and Petersson, Christoffer and {\AA}str{\"o}m, Kalle} and Svensson, Lennart,
  year = {2024},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
}

Recommend Projects

  • React photo React

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

  • Vue.js photo Vue.js

    ๐Ÿ–– 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. ๐Ÿ“Š๐Ÿ“ˆ๐ŸŽ‰

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 โค๏ธ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.