6K resolution is only possible with fish-eye
A few weeks ago we released projection meshes for the ZCam K2 Pro and K1 Pro cameras, based on the lens profiles for each model. We release videos in the original K2 Pro sensor resolution: 2900x2772 in 200º FOV, producing a 5800x2772 stereo video file that can accurately be described as 6K resolution. This resolution is only possible with a lens profile projection mesh that completely fits the sensor image. For equirectangular projection, this is cut down to 2700x2700 resulting in 5400x2700 180º FOV.
The circle in the image below shows how much extra visual footage we get with fish-eye projection. That 10º extra from each view incredibly matches the natural human FOV, meaning you won’t lose immersion through natural movements of your head. It’s also worth noting that pixel density around the edges is higher than in the center. Another advantage of fish-eye is that fewer renderings are required in the workflow process, meaning the image quality is closer to the original raw files. And processing performance of fish-eye is significantly better than with equirectangular, resulting in your GPU getting less hot and leading to longer battery life on mobile headsets.
However, some producers are creating ‘6000x3000’ (“6K”) or ‘7200x3600’ (“7K”) equirectangular videos from 5400x2700 footage that doesn’t represent the actual sensor image. It’s an upscaled image boosted with junk pixels that have no value for the viewer, but using “6K”/“7K” as a selling point. Things get even worse, because the 5K version of the video, that would otherwise be the proper resolution for videos from those cameras, are instead created by downscaling from the upscaled “6K”/“7K” files, resulting in a loss of quality.
You can compare fish-eye footage to equirectangular here: Lauren Phillips in Hard Sell (fish-eye) VS Lauren Phillips in Hard Sell (equirectangular). All of the SLR fish-eye videos can be found at https://www.sexlikereal.com/tags/fisheye-vr. We will happily provide our fish-eye mastering techniques for all interested producers. These will be released publicly soon.
We tag our 5400px equirectangular videos as “5K”. Videos submitted by other producers as “6K” are tagged as 6K. We’re unsure of the best way to deal with this issue. Ideally, we would like every VR video producer to render their videos at the original sensor resolution and not to upscale.
Commentators have mentioned the useful impact of AI or ML upscaling. Indeed this is a great technique and you can find many examples on Youtube. None of the “6K”/“7K” producers mentioned above have confirmed the use of AI or any similar tool. Video upscaling is a basic feature of any video editing software and does not enhance image quality. You end up with more bandwidth, longer download times and your headset or desktop computer heat up for no reason.
It’s also worth noting that there are no known AI upscaling techniques for VR videos that won’t result in a broken convergence plane. Simply put, AI would have to also respect stereoscopic disparity between the left and right eye and the tech is simply not there yet. And once this tech does emerge it would be right and proper to tag it as “AI upscaled”.
Please let us know your thoughts on this issue. The images below show the differences between the part of sensor image that is used for equirectangular and fish eye projections.
Update. Around the time of this blog post CzechVR released 8K video. It's the only known exception with true resolition. It's believed to be natively recorded in 8K 30FPS and interpolated to 60FPS. It's also worth noting that the same resolution from different sensors has different visual perception due to other sensor specs like 8-bit color (K1 Pro) vs 10-bit color (K2 Pro).