Awesome

PHASE: Demographic Annotations on the GCC Dataset

This is a repository for PHASE, a set of annotations to study demographic bias on uncurated text-image datasets. PHASE (Perceived Human Annotations for Social Evaluation) have been annotated with demographic and contextual attributes on images from the Google Conceptual Captions dataset.

PHASE is described in the paper "Uncurated Image-Text Datasets: Shedding Light on Demographic Bias" by Noa Garcia, Yusuke Hirota, Yankun Wu, and Yuta Nakashima.

News

• May. 2024: Code for CLIP embedding evaluation has released.
• Mar. 2023: PHASE has been selected as a highlight paper at CVPR 2023.
• Feb. 2023: PHASE has been accepted at CVPR 2023.

Collection process

For a subset of the GCC images:

1. We detect regions with people with YOLOv5.
2. We filter the regions to discard missdections.
3. Annotators annotate 4 demographic and 2 contextual attibutes per region.
4. Each attribute is annotated by 3 different annotators.

Download

Download images from here.

Download annotations from here. The zip file contains the following files:

• All the annotations (3 annotations per region-attribute):

  • phase_gcc_val_all_20221101.json
  • phase_gcc_train_all_20221101.json

• Region-level annotations (1 annotation per region-attribute after majority voting):

  • phase_gcc_val_regions_20221101.json
  • phase_gcc_train_regions_20221101.json

• Annotators information:

  • annotators.csv

Annotators statistics:

CLIP embedding evaluation

1. Extract CLIP embedding

python src/extract_clip_feature_phase.py --data_root <directory of the val annotations and images>

Please download the images and place the phase_images folder in the data_root.

2. Evaluation

For each image, we rank the captions according to the cosine similarity between their embeddings, and then compute accuracy:

python src/phase_clip_evaluation.py --data_root <directory of the val annotations>

Important information

Intended uses

The dataset can only be used for research purpose. No commercial applications are allowed. Annotators can revoke their consent to share their data at any point by contacting us.

Reference

If you find PHASE useful, please cite our research paper:

@InProceedings{garcia2023uncurated,
   author    = {Noa Garcia and Yusuke Hirota and Yankun Wu and Yuta Nakashima},
   title     = {Uncurated Image-Text Datasets: Shedding Light on Demographic Bias},
   booktitle = {CVPR},
   year      = {2023},
}