Downsampling 2x2x2?

[seung-lab]

Downsampling 2x2x2?

[william-silversmith]

Hi Albert,

I just updated the documentation. We recently added a "factor" option that lets you specify (2,2,1), (2,2,2), etc. Note that (2,2,1) and (2,2,2) are treated specially in https://github.com/seung-lab/tinybrain/ in that pains are taken to avoid integer truncation, give high performance, and in some cases, support the sparse flag.

Resolved in e07297c

[albertplus007]

Another question, if I use the google segmentation, such as 4x4x40nm resolution, and use this function to downsample the dateset to 8x8x40nm, does the same result of mine (use 4x4x40nm to get 8x8x40nm) as the dataset of google segmentation on 8x8x40nm?

[william-silversmith]

It's possible, but I doubt it unless they are also downsampling using the mode of 2x2 instead of striding. If they are using the mode, they are probably also doing it "correctly" so that the mode of mip 2 to N is based on mip 0, whereas we compute it recursively. One day I might fix this.

[albertplus007]

Ok, I see in low resolution there are lots of neurons broken, but they are well preserved at high resolution, but you know high resolution result data volume is too large, so i want to downsample the high resolution to low resolution

[albertplus007]

I am still confusing that using this library to downsample the dataset, can we maintain the dataset continuity?
How many times downsampling will this problem occur? In low resolution there are lots of neurons broken, but they are well preserved at high resolution.

[william-silversmith]

Hi Albert, Generally things seem to work ok at 32x32x40 and still work but neurites are very thin at 64x64x40. I wouldn't go higher than that for any kind of analysis that requires intact neurites.

…

On Mon, Aug 24, 2020, 9:47 PM albertplus007 ***@***.***> wrote: I am still confusing that using this library to downsample the dataset, can we maintain the dataset continuity? How many times downsampling will this problem occur? In low resolution there are lots of neurons broken, but they are well preserved at high resolution. — You are receiving this because you were assigned. Reply to this email directly, view it on GitHub <#74 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AATGQSOBZCMMKIZ5QVZGMVTSCMJ33ANCNFSM4QJBQXYQ> .

[william-silversmith]

I'm gonna close this issue as I think the question has been answered. Reopen or open a new issue if you think you need to! Happy downsampling. :)