EscVM / OIDv4_ToolKit

Download and visualize single or multiple classes from the huge Open Images v4 dataset
GNU General Public License v3.0
809 stars 635 forks source link

Support for Filtering Image-Level Labels #8

Closed mylyu closed 5 years ago

mylyu commented 6 years ago

Hi, thanks for the great toolkit. Is it possible to download images based on image-level labels (19,995 classes), rather than the 600 boxable classes only? I have downloaded the csv needed.

mylyu commented 6 years ago

I think I have found the solution. Simply modify main.py to be: name_file_class = 'class-descriptions.csv' file_list = ['train-annotations-human-imagelabels.csv', 'validation-annotations-human-imagelabels.csv', 'test-annotations-human-imagelabels.csv']

————Update———— Simply doing so works but cannot filter out images whose "Confidence" values are 0, unless you manually generate a CSV for this.

keldrom commented 6 years ago

@mylyu thx for your support :) Your request is the next evolution of our toolkit. I think I can work on it during this week. Stay tuned ;)

mylyu commented 6 years ago

@mylyu thx for your support :) Your request is the next evolution of our toolkit. I think I can work on it during this week. Stay tuned ;)

That's great.

keldrom commented 6 years ago

I'm sorry but this period is dark and full of terrors. I will produce the new release of the toolkit ASAP.

zaichang commented 5 years ago

@keldrom I'm also trying to use the toolkit to download images for a classifier rather than object detection. By modifying main.py I was able to download images from the image-level label classes (though with many urls that resulted in 404). Did you ever find a solution to filtering images with a certain confidence?

@mylyu Are you taking pull requests for this library? If so, maybe we can help with increasing it's capabilities

mylyu commented 5 years ago

@keldrom I'm also trying to use the toolkit to download images for a classifier rather than object detection. By modifying main.py I was able to download images from the image-level label classes (though with many urls that resulted in 404). Did you ever find a solution to filtering images with a certain confidence?

@mylyu Are you taking pull requests for this library? If so, maybe we can help with increasing it's capabilities

@zaichang I had a quick and dirty implementation for downloading image-level labels, but unfortunately I don't have time to clear things up and make a pull request.

keldrom commented 5 years ago

I will try to work on it this week, if I will enough free time. If you can @zaichang send to me your draft and I will use it as starting point :)

zaichang commented 5 years ago

I cobbled it together, but I can spend some time this weekend to clean it up for a pull request. Works for you? (Nice Marvin icon btw)

keldrom commented 5 years ago

@zaichang don't worry. I'm working on the implementation :)

keldrom commented 5 years ago

@zaichang @mylyu https://github.com/EscVM/OIDv4_ToolKit#download-images-from-image-level-labels-dataset

It's done! :D

I will close this issue. For any other problem don't hesitate and contact us.