Video support possibility?

[raisingdibar]

Considering the fact that video has become the best way for individuals to easily document dangerous situations (not having to focus on a subject, audio helps accountability, etc) and the fact that video is the majority of non-professional content emerging from these public spaces, has anyone considered trying to apply the existing processes to video?

I know images are the basis for video, so getting the image-processing tools ironed out is definitely a must.

I have a very basic understanding of image processing myself, so please chime in if I'm missing something that makes video unfeasible; but something tells me that there must be a way to track faces (either using ML, or gesture interactions) across the timeline of a video and maintain blur by blurring each frame according to where a face is at during that frame.

Thoughts?

[everestpipkin]

Hi! Glad to have you hop on. I've done a few tests for video but haven't had time to implement it, as it is likely almost as large of a task as the whole tool up to this point. but if you want to take on the task of video I'd fully encourage you to make a PR.

IMO facetracking is both unreliable and a little bit too heavy on mobile architecture, but the 'tap mode' I added a few days ago is an attempt to do a quicker, tap-based blur for crowds. My sketch for video processing would mean recording the tap-blur location each frame and compositing them into the video loop. My guess is this should cover most bases - the trouble will be that if there are 100s of faces in a long video, it'll take a lot of long loops to cover everyone. This could be partly addressed by doing multi-touch handling, so you can blur 5~ points at once. Still not ideal but likely still faster and more reliable than facetracking on an old phone.

[raisingdibar]

I was thinking about trying to engineer something akin to creating vector (like the shape-creating curves you'd use in Photoshop / Illustrator, for example) that could be mapped via touch against time. Then you could just have another vector for each face. But the UX of this idea sounds wack so I need to think more on it. 😄