Simplified version of face-parsing.PyTorch.
We changed the output from a 19-channel tensor to a 4-channel tensor, including Full-mask, Head-mask, Face-mask, and Inner-face-mask.
# to test a single image:
python scripts/test.py
# to compare with the original model:
python scripts/comparison.py
1st row: sample images
2-5th row: facial masks obtained from the original model.
6-9th row: facial masks obtained from our model.
The original model outputs 19-channel tensor [B, 19, H, W] and each index is matched to one of facial components as below
| index | 0 | 1 | 2 | 3 | 4 | 5 |
|---|---|---|---|---|---|---|
| component | background | skin | left_brow | right_brow | left_eye | right_eye |
| index | 6 | 7 | 8 | 9 | 10 | 11 |
|---|---|---|---|---|---|---|
| component | eye_glasses | left_ear | right_ear | ear_ring | nose | mouth |
| index | 12 | 13 | 14 | 15 | 16 | 17 | 18 |
|---|---|---|---|---|---|---|---|
| component | up_lip | down_lip | neck | neckless | cloth | hair | hat |
Our model outputs 4-channel tensor [B, 4, H, W], and each channel constructs a facial mask itself. Note that the architecture of the BiSeNet is slightly modified in our code.
- idx #0: Full Mask (Merged mask including index #1~#19)
- idx #1: Head Mask (Full Mask - (neck + neckless + clothes + hat))
- idx #2: Face Mask (Head Mask - hair)
- idx #3: Inner-Face Mask (Face Mask - (left_ear + right_ear + ear_ring))
Yukyeong Lee | yukyeongleee@gmail.com
Wonjong Ryu | 1zong2@innerverz.com