linyicheng1/leanring_triple_poses

Learning to solve for triple camera poses

learning_triple_poses

Based on CVPR2022 best paper Learning to Solve Hard Minimal Problems implements solving the relative position of three images.

Note The correctness of the pose solving is lower than the results in the paper.

We think the main reasons for this are:

• Randomised feature point selection
• Generalisation of training results

1. Build

1.1 dependencies

• Eigen
• OpenCV

1.2 build

clone this project

git clone https://github.com/linyicheng1/leanring_triple_poses.git

build project

mkdir build & cd build 

cmake .. & make -j

2. Usage

2.1 Extract feature points

./build/sift <img1_dir> <img2_dir> <img3_dir> <output.txt>

For example:

./build/sift data/000570.png data/000575.png data/000580.png data/matches.txt

After the cmd you will see the match results

2.2 Solve poses

./build/learning_triple_poses <model_path> <set_path> <data_path>

For example:

./build/learning_triple_poses ./model ./model/trainParam.txt data/matches.txt

After the cmd you will see some poses output, and the groundtruth is:

t01: 
-0.0116375
-0.022658
 0.999676
t02: 
-0.0224208
-0.0452998
 1.99962

You can find the correct result in the output poses, if not you need to run the programme again.