Method and system for automatic speaker framing in video applications
For video applications, a method for dynamically switching from a current ROI to a target ROI is disclosed, wherein the target ROI only includes active speakers. Advantageously, if the target ROI crops a non-speaker, then the target ROI is expanded to include said non-speaker. Transitioning from the current ROI to the target ROI may be achieved based on a cutover transition technique, or a smooth transition technique. The cutover transition technique achieves the change from the current arrived to the target ROI in a single interval, whereas the smooth transition technique achieves the change over a number of intervals, wherein a percentage of the total change required is allocated to each interval. A system for implementing the above method is also disclosed.
1. A method for framing video, comprising:
associating a current region-of-interest (ROI) corresponding to video of a scene being imaged;
detecting audio signals associated with the scene;
determining active speakers in the scene based on the detected audio signals;
performing a reframing operation comprising dynamically calculating a target region of interest (ROI), wherein
the target ROI is generated based on coordinates associated with a bounding box calculated for the determined active speakers in the scene, and
the target ROI is expanded to include a non-speaker when the target ROI crops the non-speaker;
calculating a degree of overlap between the current ROI and the target ROI, wherein the calculating the degree of overlap between the current ROI of the target ROI comprises determining an intersection over union (IoU) between the current ROI and the target ROI; and
based on the calculated degree of overlap, transitioning to the target ROI from the current ROI using one of a cutover transition technique and a smooth transition technique.
2. The method of claim 1 , wherein said transitioning comprises executing the cutover transition technique if the degree of overlap between the current ROI and the target ROI is smaller than a threshold and executing the smooth transition technique if said degree of overlap is greater than said threshold.
3. The method of claim 1 , wherein the smooth transition technique comprises:
calculating the total change required to transition from the current ROI to the target ROI;
determining a number of intervals over which to achieve said transition; and
allocating said total change to each interval.
4. The method of claim 3 , wherein a location of the total change across the intervals is non-linear.
5. The method of claim 3 , wherein the cutover transition technique comprises allocating the total change to a single interval.
6. The method of claim 1 , wherein each bounding box is calculated based on convolutional neural network (CNN) features.
7. A system, comprising:
a camera for capturing video;
a microphone; and
a virtual director module configured to perform a method for framing video captured by the camera, the method comprising:
associating a current region-of-interest (ROI) corresponding to video of a scene being imaged;
detecting audio signals associated with the scene, wherein the audio signals are detected by the microphone;
determining active speakers in the scene based on the detected audio signals;
performing a reframing operation comprising dynamically calculating a target region of interest (ROI), wherein
the target ROI is generated based on coordinates associated with a bounding box calculated for the determined active speakers in the scene, and
the target ROI is expanded to include a non-speaker when the target ROI crops the non-speaker;
calculating a degree of overlap between the current ROI and the target ROI, wherein the calculating the degree of overlap between the current ROI of the target ROI comprises determining an intersection over union (IoU) between the current ROI and the target ROI; and
based on the calculated degree of overlap, transitioning to the target ROI from the current ROI using one of a cutover transition technique and a smooth transition technique.
8. The system of claim 7 , wherein said transitioning comprises executing the cutover transition technique if the degree of overlap between the current ROI and the target ROI is smaller than a threshold and executing the smooth transition technique if said degree of overlap is greater than said threshold.
9. The system of claim 7 , wherein the smooth transition technique comprises:
calculating the total change required to transition from the current ROI to the target ROI;
determining a number of intervals over which to achieve said transition; and
allocating said total change to each interval.
10. The system of claim 9 , wherein a location of the total change across the intervals is non-linear.
11. The system of claim 9 , wherein the cutover transition technique comprises allocating the total change to a single interval.
12. The system of claim 7 , wherein each bounding box is calculated based on convolutional neural network (CNN) features.
13. The method of claim 1 , wherein the target ROI is generated further based on a size of the bounding box and movements of the bounding box.