When clicked, tracking state is reset to the initial condition; zero crop, reset internal states of the integrators. (This is not a property.)
Deprecated Preset will be provided through a dock. The preset group in the property dialog will be removed in the future release.
This combo-box sets the name of the preset to be loaded or saved.
To load the preset, select the preset from Preset combo-box and click Load preset button.
To save current propeties as a preset, type preset name in Preset combo-box and click Save preset button.
To delete awn existing preset, select the preset from Preset combo-box and click Delete preset button.
If you want to save only tracking parameters (Upsize recognized face and Tracking target location), enable only this check-box.
If you want to save only contol parameters (Tracking response), enable only this check-box.
These properties upsize (or downsize) the recognized face by multiple of the width or height.
The motivation is that the face recognition returns a rectangle that is smaller than the actual face.
The properties will be saved to and recalled from presets.
The frame will be scaled before sending into face detection and tracking algorithm.
Default is 2.
Larger value will reduce CPU usage but too large value will fail to detect faces.
The face detection engine requires size of the faces at least 80x80.
If you have low resolution image, it is highly recommended to set to 1.
If your resolution is much smaller, make a intermediate scene and apply face tracker filter to the scene.
- Make a blank scene.
- Put your source to the scene and expand the size of the source.
- Apply the filter to the scene.
- Put the scene to your desired scene.
These properties crop the image before sending to the face detection algorithm. The unit is pixel before scaling the image.
The properties won't affect tracking. If the face is once detected and moved out from the cropped region, the tracking will still continue.
Specify dataset for face landmark detection and enable the checkbox to calculate location and size of the face. The location is calculated by the average of all the landmark points for each face. The size is calculated by the area surrounded by the landmark points. You might need to adjust tracking target location and zoom depending on the landmark datasets.
A dataset file shape_predictor_5_face_landmarks.dat is bundled so that you can try it soon.
Original data is distributed at dlib-models.
Another model shape_predictor_68_face_landmarks.dat is ready but not bundled due to a license incompatibility.
This property sets the threshold when to stop tracking after the face is lost. During correlation tracking, scores are accumulated. When the score drops lower than the specified threshold, the tracking will be stopped.
This property set the target zoom amount in multiple of the screen size.
If set to 1.0, face size and screen size is same.
Smaller value results smaller face, i.e. less zoom.
The property will be saved to and recalled from presets.
This property set the location where the center of the face is placed.
0 indicates the center, +/-0.5 will result the center of the face is located at the edge.
The properties will be saved to and recalled from presets.
This property set the maximum of the zoom.
The property will be saved to and recalled from presets.
The tracking system has a PID control element + integrator.
This is a proportional constant. The dimension is inverse time and the unit is s-1. Larger value will result faster response.
The property will be saved to and recalled from presets.
This is a integral constant. The dimension is inverse time and the unit is s-1. Larger value results more tracking of slow move.
The property will be saved to and recalled from presets.
This is a derivative constant. The dimension is time and the unit is s. 0 will result no derivative term. Larger value will make faster tracking when the subject start to move.
The property will be saved to and recalled from presets.
This is an inverse of the cut-off frequency for the low-pass filter (LPF), which affects the derivative term of PID control element. The dimension is time and the unit is s. The LPF will reduce noise of face detection and small move of the subject.
The property will be saved to and recalled from presets.
This is another proportional constant that affects only the zoom factor. Smaller value will result in slower response of the zoom.
These parameters make dead bands and nonlinear bands for the error signal that goes to PID control element. The unit is a percentage of the average of source width and height. If the error signal is within the dead band, error signal is forced to zero to avoid small move to be tracked. The nonlinear band makes smooth connection from the dead band to the linear range.
The properties will be saved to and recalled from presets.
Face Tracker Source only This property specifies the source to take the frame texture.
This parameter sets output aspect ratio. The default is same as the source. You can choose or type any aspect ratio. If the aspect is set narrower than the source, the height will be taken from the source and the width will be calculated. If the aspect is set wider than the source, the width will be taken from the source and the height will be calculated.
Known issue: The bottom or right pixels might show flicker. The workaround is to set 1 for the crop properties in transform dialog.
These properties enables how the face detection and tracking works. Note that these features are automatically turned off when the source is displayed on the program of OBS Studio. You can keep enable the checkboxes and keep monitoring the detection accuracy before the scene goes to the program.
If enabled, face detection and tracking results are shown. The face detection results are displayed in blue boxes. The Tracking results are displayed in green boxes.
If enabled, whole image will be displayed and yellow box shows how cropped. This is useful to check how much margins are there around the cropped area.
If enabled, debugging properties listed above are effective even if the source is displayed on the program. This will be useful to make a demonstration of face-tracker itself.
Not available for released version
Save internal calculation into the specified file for each.
This option is not available without building with ENABLE_DEBUG_DATA
but still can be set through obs-websocket or manually editing the scene file to add a text property with a file name to be written.
To disable it back, remove the property or set zero-length text.
Property name: debug_data_tracker
The data contains time in second, 3 coordinates (X, Y, Z), and score of the correlation tracker. The X and Y coordinates are the center of the face. The Z coordinate is a square-root of the area. Sometimes multiple correlation trackers run at the same time. In that case, multiple lines are written at the same timing.
Property name: debug_data_error
The data contains time in second, 3 coordinates (X, Y, Z). The calculated error is the adjusted measure with current resolution, the cropped region when the frame was rendered, and user-specified tracking target. 0-value indicates the face is well aligned and positive or negative value indicates the cropped region need to be moved.
Property name: debug_data_control
The data contains time in second, 3 coordinates (X, Y, Z). The X and Y coordinates are the center of the cropped region. The Z coordinate is a square-root of the area of the cropped region.