Better Performance - Look for other algos for core.

amar-laksh / focusTV

OPENCV and Python can come together to help you binge-watch with this??!!!

GNU General Public License v2.0

8 stars 0 forks source link

Better Performance - Look for other algos for core. #6

Open Ridhwanluthra opened 6 years ago

Ridhwanluthra commented 6 years ago

So, I have been trying to use focusTV lately whenever I watch something. A couple of challenges that I believe should be dealt with:

We need to be able to ID the face of the person using it (probably no need for face recognition, just tracking should be able to solve it.)
Face tracking is also essential for reducing false pauses.
Maybe we need to go for a deep net(this may be a performance issue but with a shallow YOLO based net I believe we can get great results) or HOG based detector, due to the obvious problems with HAAR it sometimes would not provide a smooth experience such as pause with minor head shifts.

amar-laksh commented 6 years ago

Okay so answering your questions:

IDing the person seems redundant, there is no advantage to it. We just need to pause whenever no one is looking. Plus if you are not using recognition to ID that's just detection.
Face tracking is being done right now but yes we do need to better it. Maybe we could use a deep net approach to reduce false positives.
Yes I do agree we should totally head for a deep net approach.

So have a look at other approaches other than HAAR. I am gonna rename the issue also.

Nice issue, keep at it!

vaibhavk97 commented 6 years ago

Totally agree with the idea to use deep nets with YOLO as mentioned or maybe MOBILENETS for more seamless experience and faster performance. Suggestion:

Instead of tracking presence of face i think it would be better to track eyes , this would give a better experience in terms of actually watching the video.

Ridhwanluthra commented 6 years ago

i could not find the code where tracking is being done. If we use something like the KLT algorithm we can also help with https://github.com/amar-laksh/focusTV/issues/1 as it would mean we don't have to run the deep net or haar or whatever all the time but only when we lose a face in tracking. The tracking will be running 90% of the time.

Given our constraints we should

find a way to benchmark
implement tracking and test the results
compare performance and accuracy to haar
then test other face detectors
compare and iterate