COFW stands for Caltech Occluded Faces in the Wild, which is a face dataset based on real-world conditions. The dataset contains 1,007 faces and a total of 1,852 images, including 1,345 training images and 507 test images.

Faces show a lot of variation in shape and occlusion due to different poses, expressions, use of accessories like sunglasses and hats, and interactions with objects (e.g., food, hands, microphones, etc.). All images are manually annotated using the same 29 landmarks as in the LFPW dataset. The landmark locations and their occlusion/unocclusion status are also annotated.