UMDFaces Dataset

Overview

UMDFaces is a face dataset divided into two parts:
  • Still Images - 367,888 face annotations for 8,277 subjects.
  • Video Frames - Over 3.7 million annotated video frames from over 22,000 videos of 3100 subjects.

Part 1 - Still Images

The dataset contains 367,888 face annotations for 8,277 subjects divided into 3 batches. We provide human curated bounding boxes for faces. We also provide the estimated pose (yaw, pitch, and roll), locations of twenty-one keypoints, and gender information generated by a pre-trained neural network.
In addition, we also release a new face verification test protocol based on batch 3.

Part 2 - Video Frames

The second part contains 3,735,476 annotated video frames extracted from a total of 22,075 for 3,107 subjects. Again, we also provide the estimated pose (yaw, pitch, and roll), locations of twenty-one keypoints, and gender information generated by a pre-trained neural network.

Download

Before proceeding to download, please read the license carefully. The complete download details and instructions can be found in the release document. Also, please read the Readme.

Specifically, part 1 of the dataset can be downloaded from the following links: The face verification test protocol based on batch 3 of the dataset can be found in the following links. Each file is a list of 100,000 pairs of images with the corresponding label defining whether the two images belong to the same person (1 for same and -1 for different). (Please refer to the release document and our paper for more details.)
You can also download the caffe model and demo code for the fiducial keypoint detection.

The video frames for part 2 of the dataset can be downloaded from here (195GB). The corresponding bounding box annotations, keypoint locations, pose, and gender information can be found in this file.
If you want to download the corresponding videos (1.2TB), please contact Ankan Bansal.

Errata

Please note that this latest release of the dataset corrects some mistakes (subjects repeated with different names) in the older version of the dataset. For continuity purposes, the older version of the dataset can still be found on the original download link.

References

If you use our dataset or model, please cite our papers:
  • Ankan Bansal, Anirudh Nanduri, Carlos D Castillo, Rajeev Ranjan, and Rama Chellappa, UMDFaces: An Annotated Face Dataset for Training Deep Networks, Arxiv preprint, 2016.
    Bibtex entry:
    			@article{bansal2016umdfaces,
    			title={UMDFaces: An Annotated Face Dataset for Training Deep Networks},
    			author={Bansal, Ankan and Nanduri, Anirudh and Castillo, Carlos D and Ranjan, Rajeev and Chellappa, Rama}
    			journal={arXiv preprint arXiv:1611.01484v2},
    			year={2016}
    			}
    			
  • Ankan Bansal, Carlos Castillo, Rajeev Ranjan, and Rama Chellappa, The Do's and Don'ts for CNN-based Face Verification, Arxiv preprint, 2017.
    Bibtex entry:
    			@article{bansal2017dosanddonts,
    			title = {The Do's and Don'ts for CNN-based Face Verification},
    			author = {Bansal, Ankan and Castillo, Carlos and Ranjan, Rajeev and Chellappa, Rama},
    			journal = {arXiv preprint arXiv:1705.07426},
    			year = {2017}
    			}
                

Change History

  1. May 23rd, 2017: Duplicate subjects removed and new video frame dataset added.
  2. November 21st, 2016: Demo code for fiducial detector added.
  3. November 7th, 2016: New web site.




Last Modified: May 23rd, 2017. Please direct comments to Ankan Bansal