4 ms·
They didn't turn HDR mode on, that's the novel aspect of the research. They took SDR photos with different positions and exposure levels, and then merged them i
by jiggawatts 4y ago
They didn't turn HDR mode on, that's the novel aspect of the research. They took SDR photos with different positions and exposure levels, and then merged them into the NeRF.
Before modern sensors with huge dynamic range for a single photo, the "old school" way of getting HDR was to put your camera on a tripod, and take 5+ shots with different exposure levels. This can then be merged in software, but obviously this only works on static scenes and a static camera position.
The HDR-NeRF model lets you move the camera around, you don't need a tripod.
PS: Something I find hilarious is that the paper talks about HDR and shows "HDR-SDR" comparisons... but both pictures are SDR JPG images. It's pathetic how the PC world is pathologically unable to uplift imaging standards while the TV and Camera industries with far smaller R&D budgets are casually sailing past us into the future.
- zamadatix 4y agoBut what advantage does not turning HDR mode on give vs doing so is the question. I guess you could be trying to generate a NeRF with a really old smart phone?
- pedalpete 4y agoUse today could be in turning older non-HDR images into HDR NeRFs, but also as a foundation that somebody else will build something completely new on top of. Who knows what that may be.
- jasonwatkinspdx 4y agoRight now the initial hot industrial application for NeRF's are recovering radiance fields from ordinary video footage. You can have someone just walk around with an ordinary camera or phone, just making sure they have decent coverage. Then back in the studio after a bunch of crunching you get a 3d virtual set you can position a camera in arbitrarily. The quality of this is considerably better than previous approaches that were based on recovering 3d meshes. The work in the original post can now extend that process to capturing HDR data, which should enable better quality when compositing with other footage or rendered elements. And you don't need some special HDR video camera, just an ordinary one can sample the same thing.
- mrbungie 4y agoAFAIK other NeRF methods don't let you use HDR data, just LDR images, so you would have to pre-process your HDR photos (tone-mapping) and you wouldn't be able to control exposure afterwards as previous necessary information for doing so would not be present when inferring novel views. HDR-NeRF allows you to control both exposure and position when synthetizing new images. Your intuition is on the point though, with enough data/compute you could build a method that uses the photos with full dynamic range information from a HDR camera instead of "stitching" LDR photos of varying exposures like this method does.
- 1-6 4y agoPerhaps it's possible to recover overblown images from video for forensic purposes. I'd be interested in seeing some results.
- taneq 4y agoYour PS reminds me of the old ads for those Matrox 30bpp graphics cards, which ‘showed how good these cards looked’ compared with standard graphics cards. Which we watched, obviously, on a computer with a standard graphics card. Also the TV salesman telling me that a 100Hz panel would look jerky, and showing me a video on a 200Hz screen that allegedly illustrated the difference.
- cubefox 4y agoI thought this was essentially how "computational photography" worked, like in the Google Pixel photo app: The camera takes several pictures with different exposure, maybe even from different lenses, and an AI model merges them.
- ljlolel 4y agoIt is. Source: I built the ProHDR app that iOS copied