README.md

GLU-NET-Fundamental-matrix

Relative works:

• https://github.com/PruneTruong/GLU-Net
• https://github.com/JiawangBian/FM-Bench

We used the maps of GLU-Net and used it to find a fundamental matrix in FM-Bench.

Explanation on the selection process.

1. The idea behind our project is to use dense matching between two pictures using GLU-NET and extract best matching points as “feature points” using cycle-consistency criterion and test it using FM-benchmark

2. how to get the best points without checking any features inside the pictures?
   • we used cycle-consistency to evaluate each pixel

   

   • We took the best matches with the smallest distance and not bigger than a predefined threshold (between 2-4 pixel max distance)

   

   • Then we picked the best points according to their position and not picking other points inside the same square.

   

Then we tested our points on the FM-benchmark to check the quality of the dense matching.

Test your own dataset

1. Download the GLU-Net files(that related to the test files) and change the test_GLUNet.py file to the attached file.
2. Add your dataset as a folder to the main folder.
3. Run it.
4. put the results in the FM-Bench - Matches folder.
5. In Pipeline/pipeline_Demo, put in comment all the function calls except GeometryEstimation and run it.
6. Run Evaluation/Evaluate.

Our results

• Baseline is the result of FM-Bench on each dataset.

Dataset	recall	inliers before rate	inliers after rate	Threshold	distance
CPC	0.099	0.213729	0.225642	3	20
CPC	0.101	0.211742	0.222566	3	15
CPC	0.85	0.215266	0.226008	2	20
CPC	0.089	0.216761	0.228251	2	15
CPC Baseline	0.277	0.480910	0.673079

Dataset	recall	inliers before rate	inliers after rate	Threshold	distance
KITTI	0.894	0.946406	0.971144	4	20
KITTI	0.894	0.964502	0.977425	4	10
KITTI	0.898	0.948491	0.972625	3	20
KITTI	0.895	0.956532	0.97667	3	15
KITTI	0.892	0.964807	0.977560	2	10
KITTI	0.889	0.950389	0.974254	2	20
KITTI	0.897	0.957569	0.975031	2	15
KITTI Baseline	0.917	0.873992	0.979718

Dataset	recall	inliers before rate	inliers after rate	Threshold	distance
TUM	0.724	0.710168	0.744096	4	20
TUM	0.711	0.726273	0.749638	4	10
TUM	0.724	0.71075	0.744955	3	20
TUM	0.72	0.722516	0.751698	3	15
TUM	0.709	0.726295	0.751611	2	10
TUM	0.726	0.711742	0.743882	2	20
TUM	0.722	0.722569	0.748998	2	15
TUM Baseline	0.576	0.592074	0.752914

Dataset	recall	inliers before rate	inliers after rate	Threshold	distance
Tanks_and_Temples	0.464000	0.752426	0.799874	4	20
Tanks_and_Temples	0.379000	0.782090	0.826279	4	10
Tanks_and_Temples	0.402	0.785416	0.837061	3	20
Tanks_and_Temples	0.397	0.78209	0.824114	3	15
Tanks_and_Temples	0.334000	0.785679	0.828347	2	10
Tanks_and_Temples	0.381	0.792710	0.837533	2	20
Tanks_and_Temples	0.378	0.793245	0.835997	2	15
Tanks_and_Temples Baseline	0.681	0.53279	0.753079