Open mshafiei opened 6 years ago
The results u get do nor surprise me. Seems reasonable.El 2 may. 2018 15:44, Mohammad Shafiei notifications@github.com escribió:Hi,
What I am looking for:
I am interested in the task of image matching. My goal is to query one image from a dataset of images and get the entries which have a substantial overlap with the queried image.
What do I expect
When I query one image from the dataset, I expect to get a list of image indices, sorted by their scores which show how similar each image is to the queried image. The images with smaller variance from
or more overlap with - the queried image should have higher score.
What I do:
To this end, I use the demo_general example. There is the testDatabase function which serves for my purpose. Where an image is queried and the index of "similar" images are printed out along with their score.
I have tried this example code on Der Hass dataset.
I run the code with the following command:
demo_general orb \find $(DATASET_PATH)-iname *.JPG | sort
`
What is the result
And the output can be found in the attached file named as output.txt
The interesting part of the output is where it querying the database. An instance of a queries result is as follows (the entry if starts from zero),
Searching for Image 2. 4 results: <EntryId: 2, Score: 1> <EntryId: 53, Score: 0.379218> <EntryId: 76, Score: 0.36081> <EntryId: 60, Score: 0.358032>.
What is the problem? Looking at the images, one could observe that the list of images are very similar. For the example above, image 1 and image 60 have huge difference in the view point. In every other example, there are mostly one or two such images which are very different and one could find more similar images in the dataset by visual comparison. What is the question? Do I have the right expectation? Should I rely on the query scores to get the images matches with similar view points? Is there any way to make the scores more reliable in this scenario?
—You are receiving this because you are subscribed to this thread.Reply to this email directly, view it on GitHub, or mute the thread.
But the images with high scores should have similar view points. Isn't that correct? This is not the case when I run the code on Der Hass dataset. Moreover, the images mostly have similar scores. Even the ones capture from the opposite side of the queried image are similar to the ones captures near it.
Hi,
What am I looking for:
I am interested in the task of image matching. My goal is to query one image from a dataset of images and get the entries which have a substantial overlap with the queried image.
What do I expect When I query one image from the dataset, I expect to get a list of image indices, sorted by their scores which show how similar each image is to the queried image. The images with smaller variance from, or more overlap with, the queried image should have higher score.
What I do:
To this end, I use the _demogeneral example. There is the testDatabase function which serves for my purpose. Where an image is queried and the index of "similar" images are printed out along with their score. I have tried this example code on Der Hass dataset.
I run the code with the following command:
demo_general orb
find $(DATASET_PATH)-iname *.JPG | sort\
`What is the result And the output can be found in the attached file named as output.txt The interesting part of the output is where it querying the database. An instance of a queries result is as follows (the entry if starts from zero),
What is the problem?
Looking at the images, one could observe that images with high scores are not very similar to the queried image. For the example above, image 1 and image 60 have huge difference in the view point. In every other example, there are mostly one or two such images which are very different and one could find more similar images in the dataset by visual comparison.
What is the question? Do I have the right expectation? Should I rely on the query scores to get the images matches with similar view points? Is there any way to make the scores more reliable in this scenario?