Author Affiliations
1School of Mechatronic Engineering and Automation, Shanghai University, Shanghai 200444, China2Shanghai Key Laboratory of Intelligent Manufacturing and Robotics, Shanghai 200444, Chinashow less
Fig. 1. Structure of stacked-hourglass network
Fig. 2. Overall framework of face landmark detection method
Fig. 3. Structural diagram of depth network model
Fig. 4. Comparative experiment of face landmark detection on images with large posture changes and face partial occlusion
Fig. 5. CED of different methods for 300W competition dataset with inter-ocular normalization
Fig. 6. Face feature line heatmaps of 300W competition test set
Fig. 7. Detection results on 300W competition dataset
Fig. 8. CED for the Menpo competition dataset with face diagonal normalization
Condition | Method | Common subset | Challenging subset | Full set |
---|
Inter-pupil normalization | Method in Ref. [18] | 6.65 | 19.79 | 9.22 | Method in Ref. [19] | 5.50 | 16.78 | 7.69 | Method in Ref. [20] | 5.28 | 17.00 | 7.58 | Method in Ref. [4] | 5.60 | 15.40 | 7.52 | Method in Ref. [21] | 5.25 | 13.62 | 6.40 | Method in Ref. [22] | 4.95 | 11.98 | 6.32 | Method in Ref. [23] | 4.51 | 13.80 | 6.31 | Method in Ref. [24] | 4.73 | 9.98 | 5.76 | Method in Ref. [25] | 4.80 | 8.60 | 5.54 | Method in Ref. [26] | 4.12 | 8.35 | 4.94 | Method in Ref. [27] | 3.67 | 7.62 | 4.44 | FDL-PHR | 3.22 | 7.92 | 4.14 | Inter-ocular normalization | Method in Ref. [27] | 3.67 | 7.62 | 4.44 | Method in Ref. [6] | 3.33 | 6.99 | 4.05 | Method in Ref. [28] | 3.34 | 6.60 | 3.98 | Method in Ref. [29] | 3.34 | 6.56 | 3.97 | FDL-PHR | 3.11 | 5.71 | 3.62 |
|
Table 1. Error of face landmark detection methods on the 300W test set%
Condition | Method | N0.08 | Failure /% |
---|
Inter-ocular normalization | Method in Ref. [4] | 0.4294 | 10.89 | Method in Ref. [20] | 0.4312 | 10.45 | Method in Ref. [24] | 0.4987 | 5.08 | Method in Ref. [6] | 0.5212 | 4.21 | FDL-PHR | 0.6893 | 2.35 |
|
Table 2. N0.08 and failure rate of face landmark detection methods on the 300W full test set by inter-ocular normalization
Condition | Method | Error /% |
---|
Inter-ocular normalization | Method in Ref. [21]Method in Ref. [6]FDL-PHR | 13.8910.028.32 |
|
Table 3. Error of face landmark detection methods on face images with large posture changes and face partial occlusion by inter-ocular normalization
Condition | Method | N0.08 | Failure /% |
---|
Inter-ocular normalization | Method in Ref. [30] | 0.1955 | 38.83 | Method in Ref. [20] | 0.3235 | 17.00 | Method in Ref. [31] | 0.3281 | 13.00 | Method in Ref. [32] | 0.3497 | 12.67 | Method in Ref. [24] | 0.3981 | 12.30 | Method in Ref. [6] | 0.4532 | 6.80 | FDL-PHR | 0.5805 | 2.86 |
|
Table 4. N0.08and failure rate of facial landmark detection methods on the 300W competition dataset by inter-ocular normalization
Condition | Method | Meanerror | Standard deviation | Max error |
---|
Face diagonal normalization | Method in Ref. [33] | 0.0205 | 0.0340 | 0.9467 | Method in Ref. [34] | 0.0182 | 0.0179 | 0.4661 | Method in Ref. [35] | 0.0165 | 0.0235 | 0.9612 | Method in Ref. [36] | 0.0159 | 0.0201 | 0.6717 | Method in Ref. [37] | 0.0200 | 0.0756 | 0.7290 | Method in Ref. [38] | 0.0135 | 0.0095 | 0.5098 | Method in Ref. [29] | 0.0138 | 0.0157 | 0.6312 | Method in Ref. [39] | 0.0139 | 0.0260 | 0.9624 | Method in Ref. [9] | 0.0120 | 0.0060 | 0.1453 | FDL-PHR | 0.0199 | 0.0071 | 0.07184 |
|
Table 5. Error analysis of face landmark detection methods on the Menpo competition dataset by face diagonal normalization