← all work deep-computer-vision

Product matting

Alpha matte extraction for background removal.

Result

splitIoU
train0.9903
val0.9625
test0.9611

The finding

Notes

- U2-Net fine-tuned for 100 epochs with BCE plus soft-dice at 320x320. - Ground-truth mattes are derived by Otsu thresholding with hole filling.

Output

Same subject across the row; each column is the same stage in every project.

InputPublished alphaPublished cutoutThis run, alphaThis run, cutoutGround truth
InputPublished alphaPublished cutoutThis run, alphaThis run, cutoutGround truth
00210
InputPublished alphaPublished cutoutThis run, alphaThis run, cutoutGround truth
InputPublished alphaPublished cutoutThis run, alphaThis run, cutoutGround truth
01069
InputPublished alphaPublished cutoutThis run, alphaThis run, cutoutGround truth
InputPublished alphaPublished cutoutThis run, alphaThis run, cutoutGround truth
01132
InputPublished alphaPublished cutoutThis run, alphaThis run, cutoutGround truth
InputPublished alphaPublished cutoutThis run, alphaThis run, cutoutGround truth
01327

Elsewhere

Code and full results on GitHub

Every number is read from the file the run produced. github.com/hammasbuilds · kaggle.com/muhammadhammas13