top | item 41697586

(no title)

jbk | 1 year ago

I've not seen that, I can ask.

discuss

order

gcr|1 year ago

I'd actually appreciate that if it's not too much trouble. I hope this feedback from an interested potential user is useful for however the design winds up :-)

In case you're curious, a few more details about this application:

- We're taking videos outdoors, so depth is longer range than usual (0m-80m). Ideally we'd love to store depth maps as single channel fp16 alongside the RGB stream. It's common in other application to store depth as int16 (e.g. as 16-bit greyscale PNGs), but if units are mm, you get overflow at 65 meters.

- In addition to depth maps, it would be interesting if we could store segmentation maps (single-channel int16), stereo footage (maybe MV-HEVC could be useful for this if it lands in ffmpeg one day), and various camera metadata (e.g. per-video epipolar matrices)

- I've briefly looked into single-channel 12-bit h264 for this application, but if the standard provides for something less hacky, that's even better! :)

There's a brief overview of the data we collected on the project page: https://google-research-datasets.github.io/sanpo_dataset/ Check out the animation under "Dataset" for an executive summary of the various video streams.