File size: 1,568 Bytes
9aeed54 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 |
LaMem: A Large-Scale Dataset for Image Memorability
In this package, we include the images and their memorability scores, after accounting
for false alarms. The images are available in the 'images/' folder and the memorability
scores are available in the 'splits/' folder. In the splits folder, we provide information
for 5 splits of the data, each consisting of the train, val and test sets (e.g., train_1.txt,
val_1.txt, and test_1.txt).
In each txt file, we provide scores in the following way:
<image_name> <memorability_score>
where <image_name> has values like 00000001.jpg and <memorability_score> is a float
value ranging from 0 to 1.
Note that a particular image can have different scores in different splits because for
each split of the data, we use a random half of the participants to find the train and
validation memorability scores and the other half to find the test scores i.e., both the
images and participants for each of the splits is disjoint. This setting is the same as
that of prior work.
Please cite the following paper if you use this dataset in your work:
Understanding and Predicting Image Memorability at a Large Scale
Aditya Khosla, Akhil S. Raju, Antonio Torralba and Aude Oliva
International Conference on Computer Vision (ICCV), 2015
The bibtex file is available via the project website: http://memorability.csail.mit.edu
For any questions or comments, please contact Aditya Khosla (khosla@csail.mit.edu).
Please read the provided LICENSE file for the terms of use of this data.
Copyright (c) 2015 Aditya Khosla
|