just yaw?

[XiaXuehai]

Estimate the head pose by cnn, just get the yaw?what about roll and pitch?

[mpatacchiola]

Hi XiaXuehai

Yesterday I pushed the new code that estimate also the pitch. The pitch estimation is in the range [-45, +45] degrees. If you do a pull you should get the updated code. To make it work you have to run again the "sudo python setup.py install" command.

The reference code for the yaw estimation is here. You can easily expand that code to include the pitch. To implement the pitch and the yaw at the same time you can do something like that:

#Defining the session, object and loading the variables
sess = tf.Session() #Launch the graph in a session.
my_head_pose_estimator = CnnHeadPoseEstimator(sess) #Head pose estimation object
my_head_pose_estimator.load_pitch_variables("../../etc/tensorflow/head_pose/pitch/cnn_arch2_pitch_f2-29999")
my_head_pose_estimator.load_yaw_variables("../../etc/tensorflow/head_pose/yaw/cnn_cccdd_30k.tf")

#Here you should get from OpenCV the sub_frame containing the person's face

#Estimating the pitch and yaw angle
pitch = my_head_pose_estimator.return_pitch(sub_frame) #Evaluate the pitch angle using a CNN
 print("Estimated pitch ..... " + str(pitch[0,0,0]))
yaw = my_head_pose_estimator.return_yaw(sub_frame) #Evaluate the yaw angle using a CNN
 print("Estimated yaw ..... " + str(yaw[0,0,0]))

I will add some examples during the weekend. I can implement also the roll angle, the only problem is that the range is between [-25, +25] degrees and it could be not enough stable. Unfortunately there are not good public datasets with a wide range of poses for the roll angle. In future probably I will generate some artificial data to solve the problem.

All the best

[mpatacchiola]

I added a working example for estimating the Pitch here.

You can mix yaw and pitch estimation calling the two "load_variable" functions followed by the return_yaw and return_pitch ones, as explained here.

All the best

[XiaXuehai]

Thinks! Do you have some head pose datasets to test it? I test it by this,but the result was not good.

[mpatacchiola]

Actually that dataset is one of the most famous and I used it in my previous work. A network similar to the one used in Deepgaze achieved the best result ever reported on that dataset with an accuracy of 60.9% (pitch) and 62.3% (yaw) for the unknown subject test. Remember that for testing the library you have to give as input the sub frame containing the face. If you use that dataset "as is" the images are not cropped, you have to use the ground table to isolate the face and give it as input to Deepgaze.

All the best

[XiaXuehai]

Got it. Thinks a lot!