RPI Multi-View Face Database Instruction

This instruction provides information on how to use the images for training. Generally, all the images in the database are ready to use. Only a few things need to be noticed.

1.  The database include two files: original_classified.zip (it is splitted into 6 files, totally ~3.5G) and Background_Image.zip(~93M).

2.  The file original_classified.zip includes the cropped face images. All the face images are saved in different directories based on their poses and sources. The images are cropped based on ground truth, and ready to use for training. Be noticed that each face generates multiple training images with scaling, rotation and shifting. Such perturbation could increase the number of training data, and improve the robustness of training results.

3.  The poses of face are divided into 5 categories: frontal, half left profile, full left profile, half right profile and full right profile. The face patches of different poses are saved into corresponding directories.

4.  The background images are included in the Background_Image.zip. It has about 900 background images. There are no such cropped patches of background images as face images for two reasons: There are two many background patches; the background patches are quite different for different training algorithms. We suggest that the user uniformly generate background patches from the images at the first beginning, then feed back the false detections for further training.

5.  Due to copyright problem, only part of our database are distributed. For frontal face detection, getting more images from FERET and FRGC will guarantee enough training data. For profile face detection, getting more images from FERET and other sources will also provide enough training data.

6.  The user can sample the training images, and choose some of them for training. The other images can be used for validation.

If users have any question for the database, please contact us:

Qiang Ji, jiq@rpi,edu or Peng Wang,