Academia.eduAcademia.edu

Figure 3 – uploaded by Dimitris Manousos

See full PDFdownloadDownload figure
Fig. 3. Comparison between detected and tracked faces obtained with the random forest and Kalman filter (first row), and the refined head pose using the ICP algorithm (second row). The depth images have been coloured in order to facilitate the visualization, the colour information is not used in the process.

Figure 3 Comparison between detected and tracked faces obtained with the random forest and Kalman filter (first row), and the refined head pose using the ICP algorithm (second row). The depth images have been coloured in order to facilitate the visualization, the colour information is not used in the process.

Related Figures (25)

Fig. 1. Illustrative representation of the Wize Mirror. On the right widgets panel with the user graphical interface, optionally this may also include: clock, weather forecast news, etc. On the left pictorial representation of different devices used for data acquisition.
Fig. 1. Illustrative representation of the Wize Mirror. On the right widgets panel with the user graphical interface, optionally this may also include: clock, weather forecast news, etc. On the left pictorial representation of different devices used for data acquisition.
Fig. 2. Processing pipeline of the Wize Mirror methods described in this paper.
Fig. 2. Processing pipeline of the Wize Mirror methods described in this paper.
Sensitivity (True positive rate, TPR) and accuracy (Positive predictive value, PPV) experiments. First row contains the different thresholds used to consider a detection true positive. If the distance between the detected nose position and the ground truth is smaller than the thresh- old, it is a true positive, otherwise it is considered a false negative. A to- tal of 607 images were processed. RF represents the method described in Fanelli et al. (2011), whereas WM represents the proposed method. Fig. 4). Furthermore, in order to keep the real-time processing con- straint at a high frame rate, only four iterations of the ICP are exe- cuted as the results showed to be suitable for the post-processing by other functionalities of the system.
Sensitivity (True positive rate, TPR) and accuracy (Positive predictive value, PPV) experiments. First row contains the different thresholds used to consider a detection true positive. If the distance between the detected nose position and the ground truth is smaller than the thresh- old, it is a true positive, otherwise it is considered a false negative. A to- tal of 607 images were processed. RF represents the method described in Fanelli et al. (2011), whereas WM represents the proposed method. Fig. 4). Furthermore, in order to keep the real-time processing con- straint at a high frame rate, only four iterations of the ICP are exe- cuted as the results showed to be suitable for the post-processing by other functionalities of the system.
Fig. 4. Example of sub-sampling for four different levels to perform the multi-resolution registration.
Fig. 4. Example of sub-sampling for four different levels to perform the multi-resolution registration.
Fig. 5. Comparison between the two different segmentation stages used for pre- processing the input depth data for subsequent 3D reconstruction. Input depth im- ages (left), depth segmentation (centre) and model segmentation (right). The colour images are only used to visualise the segmentation results. The colour information is not used in any of the segmentation methods.
Fig. 5. Comparison between the two different segmentation stages used for pre- processing the input depth data for subsequent 3D reconstruction. Input depth im- ages (left), depth segmentation (centre) and model segmentation (right). The colour images are only used to visualise the segmentation results. The colour information is not used in any of the segmentation methods.
Fig. 6. Average face models used for face segmentation, generated using Face- Gen software. Model used for the reconstruction needed in morphological analysis of cardio-metabolic risks (left). Model utilised to build the personalised mask for tracking (right).
Fig. 6. Average face models used for face segmentation, generated using Face- Gen software. Model used for the reconstruction needed in morphological analysis of cardio-metabolic risks (left). Model utilised to build the personalised mask for tracking (right).
Fig. 7. Plastic head model used for the reconstruction experiments (left), recon- structed model using the proposed method (right).
Fig. 7. Plastic head model used for the reconstruction experiments (left), recon- structed model using the proposed method (right).
Fig. 8. Comparison between 3D reconstructions obtained using the proposed method. The images on the top represent the signed distance between the two reconstructions: the current reconstruction and a reference reconstruction. The histograms (on the bottom) are calculated with the number of points belonging to the reconstructed face (63,000 points on average) and clustered depending on their signed distance (in meters) to the reference scan. The average error for all the experiments is 1.7 + 2.4 mm.
Fig. 8. Comparison between 3D reconstructions obtained using the proposed method. The images on the top represent the signed distance between the two reconstructions: the current reconstruction and a reference reconstruction. The histograms (on the bottom) are calculated with the number of points belonging to the reconstructed face (63,000 points on average) and clustered depending on their signed distance (in meters) to the reference scan. The average error for all the experiments is 1.7 + 2.4 mm.
Fig. 9. 3D geometric reconstruction results. RGB image (left), 3D reconstruction fot morphological analysis of cardio-metabolic risk (centre), and personalised mask fot face tracking (right).
Fig. 9. 3D geometric reconstruction results. RGB image (left), 3D reconstruction fot morphological analysis of cardio-metabolic risk (centre), and personalised mask fot face tracking (right).
Fig. 10. The 23 landmarks used to analyse faces from the morphological viewpoint.
Fig. 10. The 23 landmarks used to analyse faces from the morphological viewpoint.
List of linear and planar measurements which were found to correlate with waist circumference in Lee and Kim (2014). dg stands for Euclidean distance, d, horizontal (Euclidean) distance, dy vertical (Euclidean) dis- tance, and A(p;,..., Pn) is the area of the polygon formed by points Di gany Pn. Fig. 10 explains both the position and the label of the land- Fig. 11. Geodesic (left) and Euclidean (right) distance between two landmarks.
List of linear and planar measurements which were found to correlate with waist circumference in Lee and Kim (2014). dg stands for Euclidean distance, d, horizontal (Euclidean) distance, dy vertical (Euclidean) dis- tance, and A(p;,..., Pn) is the area of the polygon formed by points Di gany Pn. Fig. 10 explains both the position and the label of the land- Fig. 11. Geodesic (left) and Euclidean (right) distance between two landmarks.
Fig. 12. Two views of each curve passing through four landmarks, on a 3D face model. In the first (resp. second) row is visualized the geodesic path a (resp. b).
Fig. 12. Two views of each curve passing through four landmarks, on a 3D face model. In the first (resp. second) row is visualized the geodesic path a (resp. b).
Fig. 13. Sections given by the intersection of the 3D face mesh with equally-spaced planes perpendicular to the z-axis.
Fig. 13. Sections given by the intersection of the 3D face mesh with equally-spaced planes perpendicular to the z-axis.
Fig. 15. A visualization of the features f3, f6, and f7. Note: due to the symmetry of the face model used, f6 and f7 are equal. Table 5
Fig. 15. A visualization of the features f3, f6, and f7. Note: due to the symmetry of the face model used, f6 and f7 are equal. Table 5
Fig. 16. Scatter plots for the subjects belonging to three groups of fatness: level 1, in red, level 5, in green, and level 10, in blue) in the embedding space given by the features f1 and f2, f3 and f6, f11 and f12, Lgeod, and Lgeod,, meanLZ and meanAZ. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article).
Fig. 16. Scatter plots for the subjects belonging to three groups of fatness: level 1, in red, level 5, in green, and level 10, in blue) in the embedding space given by the features f1 and f2, f3 and f6, f11 and f12, Lgeod, and Lgeod,, meanLZ and meanAZ. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article).
Fig. 17. Graphs of a selection of the features (f3,Lgeoda, meanLZ), computed on the whole dataset; with a zoom on the 7th seed
Fig. 17. Graphs of a selection of the features (f3,Lgeoda, meanLZ), computed on the whole dataset; with a zoom on the 7th seed
Fig. 18. Selected geometric features: f3, f6, Lgeoda, meanLZ, computed on a set of 10 subjects. Results are visualised as scatter plot of each feature vs BMI (blue) and weight (red). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article).
Fig. 18. Selected geometric features: f3, f6, Lgeoda, meanLZ, computed on a set of 10 subjects. Results are visualised as scatter plot of each feature vs BMI (blue) and weight (red). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article).
Results sample for 4 subjects. For each subject BMI and weight were collected ; and some of the geometric fea- tures implemented were computed: f2, f3, f6, Lgeoda, meanLZ. studies. For extracting the blink rate, the AAM is used in order to segment the eyelid area and to mark out the eyeball perime- ter with specific landmarks (six landmark points for each eye). Then, the average distance between the two upper and lower eye- lid points as shown in Fig. 19(a) is calculated. Eye blinks can be seen as sharp negative spikes in the extracted signal as shown in Fig. 19(b).
Results sample for 4 subjects. For each subject BMI and weight were collected ; and some of the geometric fea- tures implemented were computed: f2, f3, f6, Lgeoda, meanLZ. studies. For extracting the blink rate, the AAM is used in order to segment the eyelid area and to mark out the eyeball perime- ter with specific landmarks (six landmark points for each eye). Then, the average distance between the two upper and lower eye- lid points as shown in Fig. 19(a) is calculated. Eye blinks can be seen as sharp negative spikes in the extracted signal as shown in Fig. 19(b).
Fig. 19. Eye opening average distance calculation between the two upper and lower eye-lid points (a) and variability of eye average distance and blink threshold (b).
Fig. 19. Eye opening average distance calculation between the two upper and lower eye-lid points (a) and variability of eye average distance and blink threshold (b).
Fig. 20. Spatial distribution of landmarks on human face.
Fig. 20. Spatial distribution of landmarks on human face.
Fig. 21. Landmarks distribution (a); landmarks distribution after GPA alignment (b).
Fig. 21. Landmarks distribution (a); landmarks distribution after GPA alignment (b).
Detailed performance results of the lip twitching detection algorithm. Table 7
Detailed performance results of the lip twitching detection algorithm. Table 7
Fig. 22. Flowchart of head motion algorithm.
Fig. 22. Flowchart of head motion algorithm.
Results of the head motion algorithm, the ground truth and the accuracy. ered to be the ground truth. The duration of the videos was also pre-defined in order to extract the ground truth of speed. Eight different videos were captured with horizontal and vertical move- ments at various speeds (mm/s) and distances (mm).
Results of the head motion algorithm, the ground truth and the accuracy. ered to be the ground truth. The duration of the videos was also pre-defined in order to extract the ground truth of speed. Eight different videos were captured with horizontal and vertical move- ments at various speeds (mm/s) and distances (mm).
Related topics:
Cognitive ScienceMathematicsComputer ScienceISmart Mirror
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved