Choose dtype for isolated buffer

[kushalkolar]

Choose dtype for isolated buffer

[ArjunPutcha]

Most image data from sensors I typecast as uint 16 (in firmware) - I'm pretty sure some national instruments tools I use also do the same; generally I want to assess the raw data first to see if analog filtration is going well, not do any digital signal processing, so I'd just read in the data and visualize directly (no need for any additional typecasting).

[kushalkolar]

I propose always casting to float32 for the ImageDataFeature buffer, even simple things like Gaussian filters on uint8 or uint16 images result in overflow and meaningless results.

without casting:

with casting:

I do not think we should expect users to worry about the limitations of the dtype of their raw image data if they are using ImageWidget with frame_apply functions etc. to interactively play with different types of processing 😄 So we should make casting to float32 automatic, unless there are downsides I cannot think of?

[clewis7]

Sounds good to me...I think float32 gives enough flexibility

[ArjunPutcha]

Agreed. Might still be useful to allow users to pass in dtype if necessary with default float 32. Some image data I acquire with sensor is uint16 or uint8 by default.

[kushalkolar]

I'm not sure how useful it would be to allow users to set the dtype of the internal buffer. uint isn't useful for any basic image processing. Do you have an example use case?

…

On Fri, Nov 17, 2023, 06:21 Arjun Putcha ***@***.***> wrote: Agreed. Might still be useful to allow users to pass in dtype if necessary with default float 32. Some image data I acquire with sensor is uint16 or uint8 by default. — Reply to this email directly, view it on GitHub <#369 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACHXXRHJRFDCXEAFO6XEY6DYE5CFFAVCNFSM6AAAAAA7NFBYJKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQMJWGE4TMNZTHE> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[kushalkolar]

ah ok, so in this case you would still want to use an isolated buffer but of the same type so it's 1-1 copying into the buffer.

[ArjunPutcha]

Exactly. But I haven't benchmarked how long it takes; I think if you go from uint16 to int32, there's two conversion - uint16 to in16, and int16 to int32. I'd imagine you'd need to do the same to float32, but I'm not sure if the time required to do so is ridiculously long. I can benchmark this coming week.
Also, if I were just visualizing analog data, I'd make a bunch of sensors testing different analog settings and visualize en masse