An Upgraded Version of Image-Matching-Toolbox

This repo is developed from the original toolbox of Zhou et. al. I've made several updates as follows

• Add evaluation of my image-matching method, TopicFM+
• Update homography estimation on HPatches. I added several options for the cv2.findHomography function
• Add quantization step for the evaluation of Aachen Day-Night, based on the maximum confidence of keypoints.

We only provide the evaluation of TopicFM in this code.

For more detailed description and other functions of the toolbox, please visit the original version.

Installation

Firstly, setup the experimental environment:

conda create -n immatch python=3.8
conda activate immatch
conda install pytorch==1.12.1 torchvision==0.13.1 cudatoolkit=10.2 -c pytorch
pip install jupyter, matplotlib, opencv-python==4.7.0.72
pip install pycolmap

To evaluate on Aachen Day-Night and Inloc, install COLMAP from the original website. Make sure that you install successfully by testing colmap in your terminal.

Secondly, install the dependencies of toolbox:

cd image-matching-toolbox/
git submodule update --init
# ignore this step if you don't want evaluate other methods
cd pretrained && bash download.sh && cd ..

Clone the code of TopicFM and put it into the third_party folder

cd third_party && git clone https://github.com/TruongKhang/TopicFM 
# change the folder name TopicFM --> topicfmv2
mv TopicFM topicfmv2 && cd ..

Next, download the pretrained models of TopicFM+ and put them into third_party/topicfmv2/pretrained/. This toolbox can support evaluations of two models third_party/topicfmv2/pretrained/topicfm_fast.ckpt and third_party/topicfmv2/pretrained/topicfm_plus.ckpt.

Finally, install the toolbox as follows:

python setup.py develop

Notes: when running the program, use pip install <package-name> if there are any uninstalled packages.

Evaluation of TopicFM+

All settings of the model and datasets are specified in configs/topicfmv2.yml

HPatches

python -m immatch.eval_hpatches --gpu 0 --config 'topicfmv2' --task 'both' --h_solver 'cv' --ransac_thres 6 --root_dir . --odir 'outputs/hpatches'

AAchen Day-Night v1.1

python -m immatch.eval_aachen --gpu 0 --config 'topicfmv2' --colmap colmap --benchmark_name 'aachen_v1.1'

InLoc

python -m immatch.eval_inloc --gpu 0 --config 'topicfmv2'