Adaptive threshold produces poor image on Android

[rmdvector]

Testing our application that does image capture on a Google Pixel 7.
The code is basically straight out of the examples:
` let pipeline = gm.grayscale(im_mat);

pipeline = gm.adaptiveThreshold(

  pipeline,

  DEFAULT_ADAPTIVE_BOXSIZE_VALUE,    // 5

  DEFAULT_ADAPTIVE_THRESHOLD_VALUE,  // 15

  0 // grayscale channel only

)

const grayScale = gm.tensorFrom(pipeline)

const sess = new gm.Session()

sess.init(pipeline)

sess.runOp(pipeline, 0, grayScale)

sess.destroy()`

On desktop web (Chrome 119.0.6045.123), both our application and the example have the expected results.
However, on the Pixel, the image looks very fuzzy and full of artifacts. I tried the example for your page and saw the same results.

[rmdvector]

I'm seeing this on both Mobile Chrome and Firefox

[rmdvector]

Desktop

Mobile:

[WorldThirteen]

Hi @rmdvector, there are a huge amount of factors that might impact the result, which include but are not limited to:

• camera resolution
• output resolution
• video noise, sharpness
• calculation precision available on the device GPU.

A good place to start is to prepare a sample with a single static image that runs and outputs in the same resolution on your target devices and compare it here. This might help determine if the issue is in the input source or in the calculation itself.

[rmdvector]

Thanks for the quick response! Using identical code and browsers as above, here's the original image (edited to remove app-specific chrome):

Here's the image from the desktop web:

And here's the image from the phone:

[rmdvector]

Did github drop the images? if so is there a better way to transmit them?

[WorldThirteen]

@rmdvector, do you mean the code you provided in #194 (comment)?

Weird that GitHub has removed them, please try to re-post and I'll save them for records and further processing & debugging.

[rmdvector]

[WorldThirteen]

@rmdvector, sorry for the lack of updates here, had not yet chance to take a look deeply.