You are here
DPOSE - multi-view dataset with head pose ground truth
DPOSE dataset contains multi-view sequences with head pose ground truth acquired with 16 subjects. Each sequence comes with three parts: the subject is (i) rotating in-place at the room center, (ii) rotating in-place at some random positions, or (ii) moving around freely in a room. The dataset consists of over 50000 images, recorded from four cameras with overlapping field of view. Images have resolution 1024x768 and are stored in jpeg format. Head pan, tilt and roll measurements for various poses are recorded using an accelerometer, gyro, magnetometer platform strapped onto the head using an elastic band running down from the back of the head to the chin.
The dataset contains also an additional 5-minute recording of 6 people freely interacting during a coffee-break in the lab (this additional data is without pose ground truth).
DPOSE comes with camera calibration files: for every camera, we provide all the information needed to compute the image projection (pixel coordinates) of a 3D point. C++ code implementing this projection is provided.
We also provide tracking of the subjects (ground coordinates), computed with a color based particle filter.