ImageNet (original) (raw)
Download the Object Attributes
Annotations of object attributes are freely available for download ( no signing-in required ). The attributes are annotated and verified through Amazon Mechanical Turk.
Currently, we have 25 attributes for ~400 popular synsets available. Please click here to obtain the list of synsets available. For each synset, there are 25 images annotated with the following attributes:
- Color: black, blue, brown, gray, green, orange, pink, red, violet, white, yellow
- Pattern: spotted, striped
- Shape: long, round, rectangular, square
- Texture: furry, smooth, rough, shiny, metallic, vegetation, wooden, wet
For example:
 |
|---|
Labeling procedure (for each image and each attribute):
- Rather than labeling the entire image, we use the previously collected bounding box annotations to focus on just one part of the image which contains the object of interest.
- We ask 3-4 workers to provide a binary label indicating whether the object contains the attribute or not.
- If there is consensus among the workers, we assign the corresponding positive or negative label.
- Otherwise, we label the attribute as ambiguous for this image.
This data was initially collected for
- O. Russakovsky and L. Fei-Fei, Attribute Learning in Large-scale Datasets. Proceedings of the 12th European Conference of Computer Vision (ECCV), 1st International Workshop on Parts and Attributes. 2010.
pdf | bibtex |slides |numerical results
How to download the attributes?
- We have not yet released the attributes for all synsets. To check the list of synsets with attributes, please use the API:
- <./api/imagenet.attributes.obtain%5Fsynset%5Flist.php>
You can click here to obtain the synset names.
- <./api/imagenet.attributes.obtain%5Fsynset%5Flist.php>
- You can download all the attributes available packaged in one file here. The API will return a Matlab (.mat) file. In the Matlab file, there will be a struct attrann which contains
- A list of images,
- A list of bounding boxes (bboxes), one per image. corresponding to the labeled objects. Each bounding boxes contains the fields x1, x2, y1, y2, all normalized to be between 0 and 1,
- A list of attributes,
- A matrix of labels of size (number of images) x (number of attributes). A label of 1 (or -1) indicates the presence (or absence) of the attribute, and a label of 0 indicates lack of consensus among the workers.
Please refer to the README (corrected 1/24/15) for more details about data collection. You also can download the extracted features here. Please refer to the paper for details about feature extraction.