Correct/ incorrect matches

Moderator: Computer Vision

Ondori
Mausschubser
Mausschubser
Beiträge: 89
Registriert: 28. Okt 2006 19:50

Correct/ incorrect matches

Beitrag von Ondori »

We are wondering how to determine the correct or incorrect matches without looking at them manually.
Is there a file or should we create one?

Christoph Vogel
Erstie
Erstie
Beiträge: 20
Registriert: 30. Mär 2010 17:38

Re: Correct/ incorrect matches

Beitrag von Christoph Vogel »

The images have names which correspond to each other.
You can assume that the query image with the name objN_* corresponds to the model image with the name objN_*.
This is true for all N=1..100, or for all images in the folders.

Christian A.
Neuling
Neuling
Beiträge: 2
Registriert: 20. Apr 2010 12:26

Re: Correct/ incorrect matches

Beitrag von Christian A. »

You are right, the query and model images with the same N in the filename are actually the same objects under different viewpoints.
Unless there are multiple images of toy cars for example it brings up the following question:
Is a match a correct match if and only if the two images are the same object (same N in image name) or if they belong to the same class of objects (ex. toy cars) ?

g
chr

Ondori
Mausschubser
Mausschubser
Beiträge: 89
Registriert: 28. Okt 2006 19:50

Re: Correct/ incorrect matches

Beitrag von Ondori »

The function plot_rpc only knows the distances between images (matrix D) but not which images were compared.
Should I assume that all query images from query.txt were compared to all model images from model.txt using find_best_match?
Or is there another way I'm just not thinking of at the moment?

Thx

Christoph Vogel
Erstie
Erstie
Beiträge: 20
Registriert: 30. Mär 2010 17:38

Re: Correct/ incorrect matches

Beitrag von Christoph Vogel »

@Christian A.
To keep things simple I would strongly recommend to compare object wise, since we do not have a grouping of the images into classes here.

@Ondori
The matrix D holds the distances between all possible query model combinations available. With that information and my earlier post, you should be able to compute and draw the precision recall curves for varying thresholds.

Antworten

Zurück zu „Computer Vision“