Are you trying to steal my secret sauce ??? 
Haha, just kidding, happy to discuss, it’s all in the GitHub repo actually.
For VR, we arbitrarily apply a -21° pitch (look down), and unwarp the picture using a v360 filter depending on the type of footage (equirectangular, fisheye) and the FoV.
It’s been extensively discussed and experimented on that thread.
Feel free to join the discord and reach out from there if you feel like joining forces 
(or not, you are most welcome anyway!).
I’ve actually been experimenting with a 15° pitch which works if applied to the video beforehand, but it just strangely warps the output if I apply it to only an image/frame. I’d like to not pre-process the whole video, so I’m really hoping I can figure it out when applying the filter to images. Thanks for the hint that that’s indeed the correct approach!
And I’m most definitely not competition in any way. I initially just wanted to see how AI generation works after having tried some good and many very bad AI scripts from SLR. Then I thought I might as well try creating a model and now I’m here to see how long the project remains fun 
And yes, we run yolo on the whole movie, grab all detections in a first pass, then build a persistent locked penis box which is reset only under certain circumstances (video cut, change of position, etc.). It comes with pros and cons though. Another approach was to memorize its length until we spot again the glans and reassess its size, but this does not work so well in JAV VR. And since we like JAV VR, we had to look for a different approach.
Not sure what you call pre process in that context, but here the v360 filter that allows the VR unwarping is run on the fly.
By pre-processing I meant that my filter so far didn’t work quite well when I applied it on the fly and I had to basically make a copy of the video with the filter applied. However that’s most likely just a bug on my end.
My go to apprach regarding the penis size/length was to first scan all penis boxes that were detected before the script generation begins and estimate the (maximum/unobscured) size based on all occurences. Then I’d compare that to the visible penis area. But since I’m also a JAV fan I might have to rethink that as well
Thanks for the screenshots, I’ll really have to figure out why my pitch filter is behaving so wonky. This definitely looks like the right approach.
Here’s the ffmpeg command used on that video for instance:
ffmpeg -hwaccel videotoolbox -hide_banner -nostats -loglevel error -i /Users/k00gar/Downloads/wankzvr-halloween-fast/wankzvr-halloween-house-party-cum-slinger-180_180x180_3dh_LR_segment_4065.mp4 -an -sn -vf scale=1280:-2,crop=640:640:0:0,v360=he:in_stereo=2d:output=sg:iv_fov=190:ih_fov=190:d_fov=190:v_fov=90:h_fov=90:pitch=-21:yaw=0:roll=0:w=640:h=640:interp=lanczos:reset_rot=1 -pix_fmt bgr24 -f rawvideo pipe:1
That worked perfectly fine, thank you (I used -15° pitch)! Although he should have detected the hand there as well.. one step at a time I guess
This was the version before the filter was applied correctly about an hour ago:
It actually depends on what type of footage you trained YOLO on.
Also, in that case, you do not really care about the hand 
I’ve probably marked thousands of hand bounding boxes, so he better finds them all 
Been there. Done that. Lol
I know that feeling man 
I even worked on a calibrating endpoint mode for @roa
Nice what exactly did @roa want?
Properly timed endpoints! 
Yeah that would be a milestone if you can achieve to the frame.
Thanks again @k00gar I’ve tweeked the settings a bit more and the new 600 frame batch for annotation looks sooo much better now 
Really great little piece of software, although I feel like it has a long way before it can produce usable results organically.
There are still bugs in the new GUI, but I feel this is next level versus the previous version
Will for sure test this weekend and give you feedback.
