68 keypoint annotations for COFW test data

68 keypoint annotations for COFW test data

This repository contains manually annotated 68 keypoints for COFW test data (original annotation of CFOW dataset contains 29 keypoints). The annotations are stored at "COFW68_Data/test_annotations/". For each image, there is one .mat file which contains the location of the keypoints and their visibilities. Also, a face bounding box for each image is stored at "COFW68_Data/cofw68_test_bboxes.mat". These bounding boxes are calculated using similar detection method which is used for 300-W datasets. So a landmark localization model trained on 300-W datasets can be tested on COFW68 dataset.

Download COFW Images:

 cd	COFW68_Data
 wget http://www.vision.caltech.edu/xpburgos/ICCV13/Data/COFW.zip
 unzip COFW.zip
 mv common/xpburgos/behavior/code/pose/COFW_test.mat .

To visualize the annotations execute "VisualizeAnnotations.m"

Evaluate your method

To evaluate your model, create a text file that contains the landmark localization predictions (and their visibilities). Then specify its address and name in Main.m and execute the Main file. For each image there should be 4 lines in the text file. The first line should contain the image index (in COFW_test.mat). The second, third and fouth lines should contains x, y and occlusion state of each keypoint, respectively.

If you like to share your results with others, please send the file containing landmark localization results to [email protected] and it will be added it to the "Results" directory.

Please cite this paper if you use this benchmark for your research paper:

Golnaz Ghiasi, Charless Fowlkes, "Occlusion Coherence: Detecting and Localizing Occluded Faces", arXiv:1506.08347

Issues, Questions, etc

Please contact "gghiasi @ ics.uci.edu"