Awesome

<a href="https://xuelunshen.com/gim"><img src="https://img.shields.io/badge/Project_Page-3A464E?logo=gumtree" alt='Project Page'></a> <a href="https://arxiv.org/abs/2402.11095"><img src="https://img.shields.io/badge/arXiv-2402.11095-b31b1b?logo=arxiv" alt='arxiv'></a> <a href="https://huggingface.co/spaces/xuelunshen/gim-online"><img src="https://img.shields.io/badge/%F0%9F%A4%97_Hugging_Face-Space-F0CD4B?labelColor=666EEE" alt='HuggingFace Space'></a> <a href="https://www.youtube.com/watch?v=FU_MJLD8LeY"><img src="https://img.shields.io/badge/Video-E33122?logo=Youtube" alt='Overview Video'></a> <a href="https://community.intel.com/t5/Blogs/Tech-Innovation/Artificial-Intelligence-AI/Intel-Labs-Research-Work-Receives-Spotlight-Award-at-Top-AI/post/1575985"><img src="https://img.shields.io/badge/Blog-0071C5?logo=googledocs&logoColor=white" alt='Blog'></a> <a href="https://zhuanlan.zhihu.com/p/711361901"><img src="https://img.shields.io/badge/Zhihu-1767F5?logo=zhihu&logoColor=white" alt='Blog'></a>

<a href="https://en.xmu.edu.cn"><img src="https://img.shields.io/badge/XMU-183F9D?logo=Google%20Scholar&logoColor=white" alt='Intel'></a> <a href="https://www.intel.com"><img src="https://img.shields.io/badge/Labs-0071C5?logo=intel" alt='Intel'></a> <a href="https://www.dji.com"><img src="https://img.shields.io/badge/DJI-131313?logo=DJI" alt='Intel'></a>

✅ TODO List

• ZEB: Zero-shot Evaluation Benchmark
• Video Preprocess Code
• 3D Reconstruction
• Inference code
  • gim_roma
  • gim_dkm
  • gim_loftr
  • gim_lightglue
• Training code

We are actively continuing with the remaining open-source work and appreciate everyone's attention.

🤗 Online demo

Go to Huggingface to quickly try our model online.

⚙️ Environment

I set up the running environment on a new machine using the commands listed below.

conda create -n gim python=3.9
conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch -c conda-forge
conda install xformers -c xformers
pip install albumentations==1.0.1 --no-binary=imgaug,albumentations
pip install colour-demosaicing==0.2.2
pip install pytorch-lightning==1.5.10
pip install opencv-python==4.5.3.56
pip install imagesize==1.2.0
pip install kornia==0.6.10
pip install einops==0.3.0
pip install loguru==0.5.3
pip install joblib==1.0.1
pip install yacs==0.1.8
pip install h5py==3.1.0
pip install matplotlib
pip install omegaconf
pip install triton

🔨 How to Use the GIM Series Matching Network

1. Clone the repository

git clone https://github.com/xuelunshen/gim.git
cd gim

2. Download gim_dkm model weight from Google Drive or OneDrive

3. Put it on the folder weights

4. Run the following commands

python demo.py --model gim_dkm
# or
python demo.py --model gim_loftr
# or
python demo.py --model gim_lightglue

5. The code will match a1.png and a2.png in the folder assets/demo,</br>and output a1_a2_match.png and a1_a2_warp.png.

🎞️ Video Preprocess

Get reliable pixel correspondences between video frames without 3D reconstruction

Because of some reasons, we cannot provide specific YouTube videos used for training, but I can tell you that using the keywords walk in or walk through to search on YouTube will find relevant videos. The videos used for processing need to be shot without any processing. There should be no editing, no transitions, no special effects, etc. Below, I will introduce the entire process.

1. Put the id of the YouTube video you want to process into the video_list.txt file. For example, the id of the video https://www.youtube.com/watch?v=Od-rKbC30TM is Od-rKbC30TM. Now the video_list.txt file already contains this example video. You can do nothing now and directly go to the second step.
2. Use the command chmod +x process_videos.sh to give the process_videos.sh file execution permission
3. Use the command ./process_videos.sh video_list.txt to run the video processing code
4. Use the command python -m datasets.walk.propagate video_list.txt to run the matching label propagation code
5. Use the command python -m datasets.walk.walk video_list.txt to run the visualization code

The processing results and intermediate files are located in the data/ZeroMatch folder, and the visualization results are in the dump/walk folder. If everything goes well, you should see a result similar to the image below (click to expand the image).

conda create -n gim-video python=3.8.10
conda install pytorch==1.10.1 torchvision==0.11.2 torchaudio==0.10.1 cudatoolkit=11.3 -c pytorch -c conda-forge
pip install albumentations==1.0.1 --no-binary=imgaug,albumentations
pip install pytorch-lightning==1.5.10
pip install opencv-python==4.5.3.56
pip install imagesize==1.2.0
pip install kornia==0.6.10
pip install einops==0.3.0
pip install loguru==0.5.3
pip install joblib==1.0.1
pip install yacs==0.1.8
pip install h5py==3.1.0

🏋️ Training Network

After processing the video, it's time to train the network. The training code for gim-loftr is in the train-gim-loftr branch of the repository. The training code for gim-dkm and gim-lightglue will be released later. However, adapting the video data by gim to the architecture of dkm and lightglue is actually simpler than adapting it to loftr. Therefore, we first release the training code for gim-loftr.

1. Use the command git checkout train-gim-loftr to switch to the train-gim-loftr branch
2. Use the command below to run the training code

#! /bin/bash
GPUS=8
NNODES=5
GITID=$(git rev-parse --short=8 HEAD)
MODELID=$(cat /dev/urandom | tr -dc 'a-z0-9' | fold -w 8 | head -n 1)
python -m torch.distributed.launch --nproc_per_node=gpu --nnodes=$WORLD_SIZE --node_rank $RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT --use_env train.py --num_nodes $NNODES --gpus $GPUS --max_epochs 10 --maxlen 938240 938240 938240 --lr 0.001 --min_lr 0.00005 --git $GITID --wid $MODELID --resample --img_size 840 --batch_size 1 --valid_batch_size 2

We train gim-loftr on 5 A100 nodes, with each node having 8 GPUs with 80 GB memory. The parameters WORLD_SIZE, RANK, MASTER_ADDR, MASTER_PORT are for distributed training and should be automatically obtained from the cluster environment. If you are using single machine with single GPU or multiple GPUs, you can run the training code with the command below.

python train.py --num_nodes 1 --gpus $GPUS --max_epochs 10 --maxlen 938240 938240 938240 --lr 0.001 --min_lr 0.00005 --git $GITID --wid $MODELID --resample --img_size 840 --batch_size 1 --valid_batch_size 2

🕋 3D Reconstruction

The code for 3D reconstruction in this repository is implemented based on hloc.

First, install colmap and pycolmap according to hloc's README.

Then, download the semantic-segmentation's model parameters from Google Drive or OneDrive and put the model parameters in the folder weights.

Next, create some folders. If you want to reconstruct a room in 3D, run the following command:

mkdir -p inputs/room/images

Then, put images of the room to be reconstructed in 3D into the images folder.

Finally, run the following command to perform a 3D reconstruction:

sh reconstruction.sh room gim_dkm
# or
sh reconstruction.sh room gim_lightglue

Tips:
At present, the code for 3D reconstruction defaults to pairing all images pairwise, and then performing image matching and reconstruction,
For better reconstruction results, it is recommended to modify the code according to the actual situation and adjust the paired images.

📊 ZEB: Zero-shot Evaluation Benchmark

1. Create a folder named zeb.
2. Download zip archives containing the ZEB data from the URL, put it into the zeb folder and unzip zip archives.
3. Run the following commands

The number 1 below represents the number of GPUs you want to use. If you want to use 2 GPUs, change the number 1 to 2.

sh TEST_GIM_DKM.sh 1
# or
sh TEST_GIM_LOFTR.sh 1
# or
sh TEST_GIM_LIGHTGLUE.sh 1
# or
sh TEST_ROOT_SIFT.sh 1

4. Run the command python check.py to check if everything outputs "Good".
5. Run the command python analysis.py --dir dump/zeb --wid gim_dkm --version 100h --verbose to get result.
6. Paste the ZEB result to the Excel file named zeb.xlsx.

The data in this table comes from the ZEB: <u>Zero-shot Evaluation Benchmark for Image Matching</u> proposed in the paper. This benchmark consists of 12 public datasets that cover a variety of scenes, weather conditions, and camera models, corresponding to the 12 test sequences starting from GL3 in the table.

🖼️ Poster

📌 Citation

If the paper and code from gim help your research, we kindly ask you to give a citation to our paper ❤️. Additionally, if you appreciate our work and find this repository useful, giving it a star ⭐️ would be a wonderful way to support our work. Thank you very much.

@inproceedings{
xuelun2024gim,
title={GIM: Learning Generalizable Image Matcher From Internet Videos},
author={Xuelun Shen and Zhipeng Cai and Wei Yin and Matthias Müller and Zijun Li and Kaixuan Wang and Xiaozhi Chen and Cheng Wang},
booktitle={The Twelfth International Conference on Learning Representations},
year={2024}
}

🌟 Star History

License

This repository is under the MIT License. This content/model is provided here for research purposes only. Any use beyond this is your sole responsibility and subject to your securing the necessary rights for your purpose.