HRNetV2 Anime & VTuber Landmarks: Nose Refinement
A nose-focused fine-tune of hysts/anime-face-detector-hrnetv2, intended to reduce large nose-localization errors on stylized VTuber faces.
The architecture and original 28-point landmark ordering are unchanged: HRNetV2-W18, 256 x 256 face crops, a 64 x 64 heatmap head, and native DARK subpixel decoding with flip test. The nose is landmark 23 (zero-based).
VTuber Validation
| Nose metric | Original HRNet | Fine-tuned epoch 25 |
|---|---|---|
| P95 pixel error | 8.97 | 5.63 |
| Maximum pixel error | 14.26 | 7.70 |
| Mean NME of worst 10% (3 faces) | 0.06988 | 0.03363 |
| Error > 5% of face-box diagonal | 3 / 29 | 0 / 29 |