We have made much progress since the last time we posted.
1. We have implemented shot detect using the optical flow vectors and organized the output. So when the video is finished, a folder is created that is named "moviename_out". Within this folder, two folders are created, one called "all_shots" and one called "shots". The point of the "all_shots" folder is to store all of the clips cut that are actual shot boundaries and also when the frame moves too much in one direction. When either of these happen, we cut the clip and output it to the "all_shots" directory. The "shots" folder just contains the clips that are from actual shot boundaries. This is just for convenience for the user.
Here is a picture of inside the "moviename_out" folder:
Here is a picture of inside the "shots" folder:
Here is a picture of inside the "all_shots" folder:
2. We have made some progress on detecting faces also. We found that using multiple haarcascade classifiers for face detection provides better results than using only one, as expected. We have ran the program on a few videos and found the detection of faces to be satisfactory. Note that we have only used 2 different haar classifiers, one that shows in red and one that shows in green. For the first picture and video, we only used 1 classifier, while with the next video we used 2 different classifiers. We have many more that we will test out in the future. Here are some results from the detection:
The first picture and video is just using 1 classifier.
A picture showing the detection of about 6 dancer's faces along with their body:
Here are a couple small clips that is associated with the above screenshot:
Here are 2 clips of a different video that is using 2 haar clasiffiers (one in red and other in green):
