[Bug report] Subject segmentation returning mask of different size than image

[tomasbecket]

Describe the bug
For certain input image sizes the subject segmentation routine (16.0.0-beta1) returns mask buffer of a different size (a few pixels bigger in both dimensions), other image sizes work fine.

To Reproduce
The following code snippet and below some sample log outputs:

private val segmenter = SubjectSegmentation.getClient(SubjectSegmenterOptions.Builder()
    .enableForegroundConfidenceMask()
    .build())
suspend fun getMask(image: InputImage): Bitmap? {
    val result = segmenter.process(image).await()
    val buffer = result.foregroundConfidenceMask ?: return null
    Log.d(TAG, "Image size: ${image.width}x${image.height}=${image.width * image.height} Mask buffer size: ${buffer.position()} ${buffer.remaining()}")
    ...
}
// Image size: 320x240=76800 Mask buffer size: 0 79704
// 79704 = 324 x 246 WHY?!?!?!
// Image size: 1280x960=1228800 Mask buffer size: 0 1273654
// 1273654 = 1297 x 982 WHY?!?!?!
// Image size: 385x512=197120 Mask buffer size: 0 197120
// The above log is as expected

Expected behavior
As the documentation states (https://developers.google.com/ml-kit/vision/subject-segmentation): the foreground mask and foreground bitmap are the same size as the input image

SDK Info:
play-services-mlkit-subject-segmentation:16.0.0-beta1

Smartphone:
Pixel 8 (Android SDK 35)

[SilverDestiny]

Could you provide the original problematic images so that I can try them out on my side? Thanks!

[tomasbecket]

Thanks for coming back to me, attached is the image (creating InputImage with rotation of 90 degrees), also attaching the mask buffer that we've gotten from the subject segmenter.

sample_1739265532691.bin.zip