Landmark Prediction on MAFL (test)
3.08Mean Error (%)Jakab et al. [21]
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Jakab et al. [21]Supervision type=Self-supervised, Training data=CelebA2018.08 | 3.08 | — | |
| Unsupervised Discovery of Object Landmarkssupervision=Unsupervised discovery, number of landmarks=302018.04 | 3.15 | — | |
| Ours# discovered landmarks=30, Regressor training set=CelebA2018.04 | 3.15 | — | |
| Zhang et al. [64]Supervision type=Self-supervised, Training data=CelebA2018.08 | 3.15 | — | |
| Ours# discovered landmarks=30, Regressor training set=MAFL2018.04 | 3.16 | — | |
| FAb-Net w/ curric., 3 source framesSupervision type=Self-supervised, Training data=VoxCeleb+, Curriculum learning=true, Number of source frames=32018.08 | 3.44 | — | |
| Unsupervised Discovery of Object Landmarkssupervision=Unsupervised discovery, number of landmarks=102018.04 | 3.46 | — | |
| Ours# discovered landmarks=10, Regressor training set=CelebA2018.04 | 3.46 | — | |
| Ours# discovered landmarks=10, Regressor training set=MAFL2018.04 | 3.46 | — | |
| FAb-Net w/ curric.Supervision type=Self-supervised, Training data=VoxCeleb+, Curriculum learning=true, Number of source frames=12018.08 | 3.49 | — | |
| FAb-NetSupervision type=Self-supervised, Training data=VoxCeleb+, Curriculum learning=false, Number of source frames=12018.08 | 3.78 | — | |
| Thewlis et al.# discovered landmarks=50, Regressor training set=CelebA2018.04 | 5.33 | — | |
| MTCNNsupervision=Fully supervised2018.04 | 5.39 | — | |
| MTCNN [65]Supervision type=Supervised, Training data=CelebA2018.08 | 5.39 | — | |
| Dense-DAELatent vector size=962018.06 | 5.45 | — | |
| Dense-DAELatent vector size=642018.06 | 5.5 | — | |
| DAELatent vector size=642018.06 | 5.7 | — | |
| DAELatent vector size=162018.06 | 5.71 | — | |
| Thewlis et al.# discovered landmarks=30, Regressor training set=CelebA2018.04 | 5.76 | — | |
| Thewlis et al.supervision=Unsupervised discovery, representation=dense frames2018.04 | 5.83 | — | |
| Thewlis et al.2018.06 | 5.83 | — | |
| Thewlis et al. [52]Supervision type=Self-supervised, Training data=CelebA2018.08 | 5.83 | — | |
| VGG-Face descriptor [41]Supervision type=Supervised, Training data=CelebA2018.08 | 5.92 | — | |
| DAELatent vector size=322018.06 | 5.96 | — | |
| Thewlis et al.# discovered landmarks=10, Regressor training set=CelebA2018.04 | 6.32 | — | |
| DAELatent vector size=962018.06 | 6.46 | — | |
| Thewlis et al.supervision=Unsupervised discovery, number of landmarks=502018.04 | 6.67 | — | |
| Thewlis et al.# discovered landmarks=50, Regressor training set=MAFL2018.04 | 6.67 | — | |
| Thewlis et al. [51]Supervision type=Self-supervised, Training data=CelebA2018.08 | 6.67 | — | |
| Dense-DAELatent vector size=162018.06 | 6.85 | — | |
| Thewlis et al.# discovered landmarks=30, Regressor training set=MAFL2018.04 | 7.15 | — | |
| TCDCNsupervision=Fully supervised2018.04 | 7.95 | — | |
| TCDCN2018.06 | 7.95 | — | |
| Thewlis et al.# discovered landmarks=10, Regressor training set=MAFL2018.04 | 7.95 | — | |
| Cascaded CNNsupervision=Fully supervised2018.04 | 9.73 | — | |
| DAELatent vector size=32, Estimation mode=Res (Residual)2018.06 | 9.93 | — | |
| DAELatent vector size=32, Regularization=NR (None)2018.06 | 10.24 | — | |
| CFANsupervision=Fully supervised2018.04 | 15.84 | — | |
| Jakab# Landmarks=10, Evaluation Protocol=Linear regression of unsupervised landmarks2019.03 | — | 3.19 | |
| Ours# Landmarks=10, Evaluation Protocol=Linear regression of unsupervised landmarks2019.03 | — | 3.24 | |
| Thewlis# Landmarks=10, Evaluation Protocol=Linear regression of unsupervised landmarks2019.03 | — | 6.32 | |
| Zhang# Landmarks=10, Evaluation Protocol=Linear regression of unsupervised landmarks2019.03 | — | 3.46 |