Synthetic PIE Dataset Documentation

GroundTruth

In addition to providing the datasets we have also included ground truth of the eye coordinates used in our experiments as well as the ground truth we used with the PIE dataset. The ground truth is in the format of filename/lefteye X/lefteye Y/righteye X, righteye Y. Download the groundtruth package here

Ground Truth Package Info  
Dataset Has Ground Truth
Original PIE Yes
ScreenShot Yes
Distance Some
Blur Yes
Pose Yes

3D Models

The 3D models were generated from the original PIE dataset using the Forensica software package from Animetrics, Inc. All 68 subjects were generated using two images from the lights subset, one frontal and one right profile. Examples of the images used are below. Model rendering examples can be seen in the Screen Shot section. The 3D data created by Forensica is output in wavefront object file format. The textures are in Portable Network Graphics (PNG) format. Download 3D models here

Screen Shot

The screenshot database contains all 13 camera positions from the original PIE. Refer to documentation from the pose dataset for the angles we used for OpenGL rotate function. The left picture is is example of PIE subject 4006 Camera 27(frontal, no  vertical or horizontal angle change). The right picture is PIE subject 4013 Camera 37(Vertical angle 2 , Horizontal angle 44). Download the PIE screenshot dataset here.

Screen Shot Dataset Info
Subjects 68
Pose Angles 13
Distance N/A
Ground Truth Yes
Notes:
Camera 27 frontal (V0, H0)  Camera 37 (V2, H44)

Frontal Long Distance

Blur Dataset Info
Subjects 68
Pose Angles 1
Distance PIE-3D-20100210B is 81 meters, all other datsets are 214 meters
Ground Truth Yes
Notes:

Dataset PIE-3D-20100210B missing subjects 4058,4068,4069.

PIE-3D-20100224C missing subjects 4000, 4001.

PIE-3D-20100224D missing 4000.

This is our first attempt at creating a long distance dataset with the Synthetic PIE Photoheads. All of them are of heads with no pose varation. PIE-3D-20100210B is taken at 81 meters indoors. PIE-3D-20100224A,PIE-3D-20100224B,PIE-3D-20100224C and PIE-3D-20100224D are taken 214 meters away outside. Download Frontal Distance datasets here.

Blur

Using a shaker table on the leg of the tripod we were able to create a motion blur set. Below is an example of subject 4006 screenshot, re-image at 81 meters, and re-imaged at 214 meters with and without motion blur. Download the motion blur dataset here.

Blur Dataset Info
Subjects 68
Pose Angles 1
Distance 214 meters
Ground Truth Yes
Notes: Camera 22 is missing subject 4069
Screen Shot Hallway 81 Meters

214 Meters Outdoors No Motion Blur 214 Meters Outdoors With Motion Blur

Pose

The pose dataset is a recreation of all the 13 cameras of original PIE but at distance. This is simulated by moving the head in the rendering program to the equivlent orientation. All the angles are pulled directly from PIE’s documentation of the camera angles except for camera 07. Their document angle did not show any visible change to rotation of the head from camera 27. Instead for this one we estimated by comparing the real image and the synthetic render. Below is an example of Subject 4006 re-imaged at a distance of 214 Meters with the vertiacal and horizontal rotatin angles listed next to each camera. Download the PIE Pose here.

Pose Dataset Info
Subjects 68
Pose Angles 13
Distance 214 Meters
Ground Truth Yes
Notes: Camera 22 is missing subject 4069

–>