coallaoh/ImageNet-AB

General Information

Title: ImageNet-AB

Description: ImageNet-AB is an extended version of the ImageNet-1K training set, enriched with annotation byproducts (AB). In addition to the image and corresponding class labels, this dataset provides a rich history of interactions per input signal per front-end component during the annotation process. They include mouse traces, click locations, annotation times, as well as anonymised worker IDs.

Links:

• ICCV'23 Paper
• Main Repository
• ImageNet Annotation Interface

Collection Process

Collection Details: The additional annotations for the ImageNet-AB dataset were collected using Amazon Mechanical Turk (MTurk) workers from the US region, due to the task being described in English. The task was designed as a human intelligence task (HIT), and the qualification approval rate was set at 90% to ensure the task's quality. Each HIT contained 10 pages of annotation tasks, each page having 48 candidate images. We follow the original annotation interface of ImageNet as much as possible. See GitHub repository and Paper for further information.

Annotators interact with different components in the annotation interface, using input devices. This interaction results in time-series data for mouse movements (mouseTracking) and mouse clicks (selectedRecord) for every image. The dataset also records whether the image was ultimately selected by the annotator in the 'selected' field.

Annotator Compensation: Annotators were paid 1.5 USD per HIT. The median time taken to complete each HIT was 9.0 minutes, yielding an approximate hourly wage of 10.0 USD. This wage is above the US federal minimum hourly wage. A total of 20,304 USD was paid to the MTurk annotators, with an additional 20% fee paid to Amazon.

Annotation Rejection: We rejected a HIT under the following circumstances.

• The recall rate was lower than 0.333.
• The total number of selections among 480 candidates was lower than 30.
• The annotator did not complete at least 9 out of the 10 pages of tasks.
• The annotation was not found in our database, and the secret hash code for confirming their completion was incorrect.
• In total, 1,145 out of 14,681 completed HITs (7.8%) were rejected.

Collection Time: The entire annotation collection process took place between December 18, 2021, and December 31, 2021.

Data Schema

{
  "imageID": "n01440764/n01440764_105",
  "originalImageHeight": 375,
  "originalImageWidth": 500,
  "selected": true,
  "imageHeight": 243,
  "imageWidth": 243,
  "imagePosition": {"x": 857, "y": 1976},
  "hoveredRecord": [
    {"action": "enter", "time": 1641425051},
    {"action": "leave", "time": 1641425319}
  ],
  "selectedRecord": [
    {"x": 0.540, "y": 0.473, "time": 1641425052}
  ],
  "mouseTracking": [
    {"x": 0.003, "y": 0.629, "time": 1641425051},
    {"x": 0.441, "y": 0.600, "time": 1641425052}
  ],
  "worker_id": "47DBDD543E",
  "assignment_id": "36DSNE9QZFQKOCZGAHS6R63J6E1OJL",
  "page_idx": 3
}

Usage

One could use the annotation byproducts to improve the model generalisability and robustness. This is appealing, as the annotation byproducts do not incur extra annotation costs for the annotators. For more information, refer to our ICCV'23 Paper.

Dataset Statistics

There were two annotation rounds covering 1,281,167 ImageNet1K training images. In the first round, annotators re-selected 71.8% of these images. The remaining 28.2% were re-packaged into a second batch of HITs, from which an additional 14.9% were selected. In total, 1,110,786 (86.7%) of ImageNet1K training images were re-selected, with annotation byproducts available for 1,272,225 (99.3%) of the images.

Other dataset statistics are inherited from the parent dataset, ImageNet-1K.

Ethics and Legalities

The crowdsourced annotators were fairly compensated for their time at a rate well above the U.S. federal minimum wage. In terms of data privacy, the dataset maintains the same ethical standards as the original ImageNet-1K dataset. Worker identifiers were anonymized using a non-reversible hashing function, ensuring privacy.

Our data collection has obtained IRB approval from an author’s institute. For the future collection of annotation byproducts, we note that there exist potential risks that annotation byproducts may contain annotators’ privacy. Data collectors may even attempt to leverage more private information as byproducts. We urge data collectors not to collect or exploit private information from annotators. Whenever appropriate, one must ask for the annotators’ consent.

Citation Information

Detailed citation information is to be provided.

@inproceedings{han2023iccv,
  title = {Neglected Free Lunch – Learning Image Classifiers Using Annotation Byproducts},
  author = {Han, Dongyoon and Choe, Junsuk and Chun, Seonghyeok and Chung, John Joon Young and Chang, Minsuk and Yun, Sangdoo and Song, Jean Y. and Oh, Seong Joon},
  booktitle = {International Conference on Computer Vision (ICCV)},
  year = {2023}
}

Maintenance and Updates

This section will be updated as and when there are changes or updates to the dataset.

Known Limitations

We have not been able to acquire annotation byproducts for all original ImageNet-1K dataset samples. This is because not all ImageNet-1K samples are re-selected by the annotators, potentially because of the errors in the original ImageNet-1K dataset. Given the budget constraint, we have not been able to acquire 10+ annotations per sample, as done in the original work.