1/9/2019 FEC Dataset - Google Docs
https://docs.google.com/document/d/17Nk5csaPeaBZv0730l4PhthJ8GuXa0fRaAMZ_D3wRy0/edit 1/3
Facial expression comparison dataset
This dataset consists of face image triplets along with annotations that specify which two faces
in the triplet form the most similar pair in terms of facial expression. To the best of our
knowledge, there is no existing large scale face dataset with such expression comparison
annotations.
Dataset content:
The dataset is released as CSV files. Each line in a CSV file corresponds to one data sample,
which consists of three face images and annotations (by multiple human annotators) that
indicate which two faces in the triplet form the most similar pair in terms of expression.
Each face image is specified using an image url and a face bounding box (top-left and
bottom-right coordinates). The users need to download the images themselves using the
provided urls. The face regions can be cropped from the downloaded images using the provided
bounding boxes.
Note: Some of the images in this dataset may not be publicly available (if an owner removes
their images from public domain) when a user is trying to download them. In such cases,
depending on the hosting website and the tool used to download, a dummy image may get
downloaded instead of the original image. We recommend users to run a face detection
algorithm on the downloaded images and verify that a face is indeed present at the location
specified by the provided bounding box.
Each annotation is an integer from the set {1, 2, 3}. A value of 1 means the expressions on
second and third faces in the triplet are visually more similar to each other when compared to
the expression on the first face. A value of 2 means the expressions on the first and third faces
in the triplet are visually more similar to each other when compared to the expression on the
second face. A value of 3 means the expressions on the first and second faces in the triplet are
visually more similar to each other when compared to the expression on the third face. Each
human annotator who participated in the annotation process has a unique id. Most of the
samples in this dataset were annotated by six human raters. However, there are few samples
which have more than six annotations.
Each triplet in the dataset is also categorized into one of the following three types: one-class
triplet, two-class-triplet and three-class triplet. Please see the next section for the definition of
these types.
