PDF version of this entire document

List of Figures

1. Expression parameterisation in action (image from Al-Osaimi et al.)
2. The effect of varying the first (top row), second, and third parameter of a brain appearance model by $\pm2.5$ standard deviations
3. The analysis performed by a Robust Generalised PCA algorithm
4. The proposed framework for GMDS improvement
5. A closer look at the GMDS approach
6. Crude visual example of how typical PCA and GMDS relate to one another, approach-wise
7. Example face from the FRGC dataset
8. 3-DImage example from the FRGC dataset, demonstrating points on the side of the face - points which need to be removed
9. Example face with holes remaining in the data
10. Another example
11. Same as above, different angle
12. Program steps broken down into an overview-type flowchart
13. Areplotted block diagram of the components in the IJCV paper (top) and our proposed extension/modifications (bottom), and the already-implemented procedures
14. Translation of the given (cropped) face applied so as to position it with the nose tip at the front and at the centre
15. A before/after overview
16. Early prototype of the GUI
17. The same GUI at a later date
18. The shape-residual extracted from two different images of different people, where the faces are aligned so as to fit a common frame of reference.
19. A sample image and a corresponding residual wrt to another (unseen) image
20. Example of what happens when the nose is incorrectly identified
21. Examples of challenging residuals that have a lot of noise
22. Process of cleaning up the residual of two images (GIP data)
23. Examples of residual images before and after outliers removal
24. Examples of faces that, given the default set of parameters, do not detect the nose correctly (with old-style cropping)
25. Improved cropping of faces that takes spatial measurements into account
26. Example of binary masks being applied to image residue
27. Examples of 4 raw residuals being reduced (notice the Z scale) using thresholds and masks
28. Cropping of GIP data shown on the top left
29. Model modes decomposed for a couple of GIP datasets (abbreviated to account for top 10 modes alone)
30. Example 3-D representation of an arbitrary face image from FRGC 2.0
31. Example of alignment and cropping of the rigid parts of another face (mind axes scale) with the result shown at the top surface and right image inside the GUI
32. Example of alignment and cropping of the rigid parts of another face with the result shown at the top surface and right image inside the GUI
33. Example of alignment and cropping of the rigid parts of another face there there is some noise and hole that pose a challenge
34. Decomposition based on sample GIP data (Pareto)
35. Decomposition based on just 4 registered faces with different expressions (Pareto)
36. Same as prior figures, but with 90 images in the set
37. Overview of an experiment dealing with expression-to-expression variation model-building
38. The masks used to crop residuals in the FRGC dataset. The left-hand one is more restrictive and selective in the sense that it omits some of the data associated with the face near the edges.
39. The top row shows images of the same subject and the bottom one is a group of hard cases (image from Huang et at.)
40. Left to right: Texture image mapped to 2-D, 3-D representation, and cropped parts (for alignment)
41. Example of an image where the face does not fit the image frame, unlike the example at the top right
42. Image residue incorrectly cropped by a binary mask
43. Examples of image residues from the FRGC datasets
44. Image residues from the FRGC datasets shown from frontal angle
45. The image shows a matrix corresponding to two things; on the left there is a top-down view of what is shown on the right. The top part shows how the 15,000 sampled (cloud)points get distributed after dimensionality reduction and the bottom part relates to the magnitude of the principal components, where the red parts show higher deviation from the mean. It is fairly smooth.
46. Principal components as a function of datapoints (log scale)
47. Point-to-point correlation along the 10th utmost principal axis
48. Score per sample point, mapped back from vectorised form to the image grid
49. The tenth principal component derived from PCA, as visualised based on the reshaped (previously vectorised) image residues
50. The tenth image residual from the Experiment 3-built EDM projected onto the EDM space's most significant component
51. Example of an evaluation experiment for photometric ICP
52. ExamExample points cloud for ICP to register
53. On the left: two faces (with binary masks cropping them for rigid parts like nose and forehead) overlaid for ICP; on the right: same from another angle
54. Examples of the faces used tor training and recognition, with neutrals on the left and smiles on the right
55. Examples of the program with the new data and methods in place
56. Pixel-wise difference between the images from our expressions set and resultant corresponding images following ICP-found translation
57. The two images that automatic detection struggles with (because of the hair)
58. The first six neutral images taken from the set and cropped by the algorithm correctly
59. The first 6 images in the set with a narrow mask used to extract and attain a neutral-to-non-neutral residue
60. Same as the previous figure, but with only 5 images. The top row shows the effect of using a broader mask and the bottom part shows the effect of applying a fixed mask and thresholds to make the data more trivially comparable.
61. ROC curves plotted for just 13 tests done on the FRGC datasets with expressions isolated
62. Preliminary test where several images (not complete set) are used to get a rough idea of what the ROC curves will look like
63. A somewhat larger test on non-neutral sets, where ICP based on PCA is used for alignment, then mean of residuals get used as a similarity measure
64. The same type of comparison with the same type of set (as in Figure ) but with a more cunning similarity measure and an example of the X data at the bottom right
65. A combined view of X, Y, Z, the image before ICP, after ICP, and the reference image
66. Difference images of the first 5 pairs taken from the same people
67. Difference images of the first 12 pairs taken from different people
68. ROC curve of the 17 images from figures and
69. The curve showing the performance for 83 pairs from false matches and 37 from true matches
70. Images ~/NIST/FRGC-2.0-dist/nd1/Fall2003range/04557d337.abs and ~/NIST/FRGC-2.0-dist/nd1/Fall2003range/04557d339.abs, where there is some detection difficulty
71. Example face-to-face comparisons
72. The ROC curves comparison on he left as linear scale and on the right log-scaled
73. The results in terms of recognition rate after widening the mask and also changing from median to mean
74. The results from full-resolution image sets and low-resolution equivalents (as seen at the top left)
75. The FP analysis of the results of a model-based approach, with the breakdown of modes shown at the top right
76. Performance comparison between an approach where the median of squared differences gets compared to mean of model changes
77. Performance of recognition when the absolute differences are gathered by their median
78. Performance of recognition when the squared differences are gathered by their means
79. Degraded performance when the compared face pairs are not ones that were used to train the PCA model
80. Comparison assessment with a large set of false pairs. The results of random pairs versus unseen pairs with expression differences.
81. The results of matching random pairs from different people and from similar people, with and without expression (based on expression models)
82. The results of comparing correct pairs to random (and false) pairs using the model-based approach. The right hand side shows the breakdown of model modes.
83. Performance measured on relatively small sets, empirically showing that coarser grids yield better recognition performance
84. Decomposition of the different modes of variation for the three cases, namely granularity levels 6x6, 8x8, and 10x10, respectively (10 pixels/voxels apart) demonstrating that despite the changes in resolution the model modes have a similar distribution and are probably inherently similar, as expected
85. Comparison between the performance of the method with smoothing applied before sampling and without any smoothing at all (which gives similar performance)
86. The decomposition of the model as a chart corresponding to Figure on the right, this time with smoothing on
87. The first few face residues following alignment with ICP (sample points being around the forehead, nose, and eyes)
88. The results of measuring the similarity by determinant of the eigenvalues of the covariance matrix and engaging in a recognition task
89. Results of an experiment where the determinant is again being explored, this time with a larger set
90. Results of an experiment where the determinant is again being explored with a comparison of the curves for 3 values of $\delta$
91. The result - in terms of performance - of varying $n$ in $\lambda_{1<i<n}$
92. The effect of changing the value of $\delta$ on the overall recognition performance
93. An 8x8 separation between points in the image (shown from two angles), with downsampling done for debugging purposes
94. A slice or subset of the data being used for ICP (on the left) and the masked face from which it is extracted (right)
95. Top: Two images taken from the same individual being compared when there is insufficient compensation for noise. Bottom: another set of such images but where smoothing is applied to reduce noise-imposed anomalies
96. On the left: the result of poor or buggy ICP (difference); on the right, an image is shown of the type of image we expect to have and also get when ICP performs well
97. Difference between the first 4 images before and after ICP (rotation and translation), with two of the first reference images shown at the bottom just for a sense of what the images at the top are derived from
98. The effect of the bug demonstrated by showing misalignment on the X axis (and to a lesser degree in Y too).
99. The new distribution of modes following the bugfix
100. The effect of perturbing the points on ICP
101. The effect of noise on ICP studied by aligning images 1-5 at the top to images 6-10 at the bottom
102. The purely median-based performance on the Spring Semester set, without ICP
103. The purely model-based (determinant) performance on the Spring Semester set, without ICP
104. Example differences between an image before and after translation (in all three dimensions)
105. On the left: the results from a test run (first 20 images) using the determinant-based objective function. The model was not constructed with translation, whereas matching did. On the right: the same but with a median-based similarity measure.
106. A top-to-bottom view of one's face (the rigid part) with corresponding translation and rotation
107. A look at some of the tweaking and debugging process of ICP, where the angle shown is pointing from underneath the nose, going towards the top
108. An example of the first image pair, visualised separately as stripes as coarse as the image sampling rate (for the model)
109. The results of a mis-constructed experiment where ICP did not work correctly and nonetheless, the model-based approach did not fail so miserably
110. Aligned and misaligned derivative difference (Y only)
111. Multi-feature experimental data (y-only derivative)
112. Y derivative (left) and X derivatives (right)
113. A couple of faces with the Y derivative on the left and the X derivative of the Y derivative (result of a bug) on rte right
114. An example Y derivative image before (left) and after (right) signal enhancement
115. The result of a very crude experiment on Fall Semester datasets, which build a PCA model of derivative differences and then perform recognition tasks on unseen faces. ROC curves are shown on the left, the composition of the model is abstracted on the right.
116. The result of a buggy code creeping into experiment as in (incorrect values were sampled). ROC curves are shown on the Left, the composition of the model is abstracted on the right.
117. The result of a correct code dealing with an experiment like in but with data from the Fall Semester
118. The effect of stress minimisation of the shape of a cat
119. Randomly chosen face sampled 10 point apart along each dimension
120. Example of almost randomly selected distances along the shapes
121. Improved selection of distances (787 vertices) and the effect of MDS reducing the stress
122. Top: original image. Bottom (from top to bottom, left to right): stress minimisation with MDS, one iteration at a time
123. A look at the cruder among ways to perform a comparison between faces
124. An exploratory look at how applying MDS to face images of the same subject depends on presupplied distances
125. Transformation from 3-D face (left) to a subset of rigid parts and then GMDS handling of the underlying surface (right)
126. Nose and eye regions from different people (FRGC 2.0) as treated by GMDS ($N=50$)
127. Nose and eye regions from different people (FRGC 2.0) as treated by GMDS when $N=100$
128. Nose and eye regions of the same person (FRGC 2.0) as treated by GMDS
129. The first pair in the set of real matches (same person in different poses)
130. An example of a problematic pair with a false signal spike (left)
131. A view of the program's front end (framework wrapper)
132. A view of the handling of image pairs and their comparison using GMDS
133. A simple visualisation of the algorithm's processing of images, by numbers
134. The correspondence problem in GMDS and an abstraction of the data by consideration of a top-down representation
135. Performance tests on very basic GMDS algorithm applied to rigid face parts
136. Examples of some of the bugs encountered and overcome while working on GMDS implementation for faces
137. Set of results for 10x10 grid sampling (GMDS)
138. Early performance measures for GMDS more properly done
139. Larger scale examples of early performance measures
140. Results from poor PCA model, obtained using GMDS
141. Model modes distribution, corresponding to Figure
142. Model modes distributions (of 10 people and 76 people), built with the proper weight, albeit with very heavy and sometimes excessive smoothing
143. Preliminary results from GMDS-based recognition with full face surface
144. A look at an alternative mask which focuses on the nose and inner eye only
145. Recognition results based on the mask from with GMDS
146. A nose-only mask, which omits areas with potential of facial hair (the examples at the centre and the left are not related)
147. The performance attained by applying GMDS just to the nose region
148. Example of the effect of ICP-induced rotation on the Voronoi cells
149. Return to the old mask with additional rotation, which does not yield better results than those at the region of 92%-98% recognition rate
150. Cheek inclusion gradually staged in for understanding of its impact on recognition performance (geodesics and PCA)
151. The stress map corresponding to the new binary mask (with 150 points for FMM)
152. Stress map and the corresponding faces (looking from beneath the nose) from which it is derived
153. An example of a false pair (different people) and a cleaned up stress map showing some interesting patterns
154. Another example of a false pair and the results of GMDS
155. Example of a bug found in the program, leading to massively false correspondence upon the same person
156. An example of acceptable matching between two poses of the same person
157. Another example of a bug found (and resolved) after it had proven problematic to recognition rates
158. A look at the problem associated with narrow faces that lead to incompatible sampling
159. General program settings used for the subsequent experiments
160. Results of a large-scale test after previous bugfixes
161. Example of a correspondence problem in a pair of images (one image on the left, another on the right). The top 4 images show the correspondence after the bugfix for one pair and the bottom 4 show the outcome of applying a fix to another pair.
162. Example of a problematic pair where hair obstruction and nose position compared to the forehead caused an issue which is now properly addressed
163. Example of a pair where the side of the face got sampled, leading to serious issues (top) before they got resolved (bottom)
164. Comparison between images of the same person, where the height of the nose relative to the cropping is causing issues
165. The images corresponding to the above example (same person, different positions)
166. Another example of a problematic example where the score borders on being seen as ``no match'' even though it is
167. Some recognition results from the above experiments, with denser sample on the right where the cheeks were also remove to test their impact on performance (little impact)
168. Smaller-scale and large-scale (right) experiments that look at how applying the methods only to the training set (many identical faces clustered together) changes the above results. It does not affect them much.
169. Standard program settings with which to run the Texas data
170. An example of GMDS applied to just a vertical slice of the data taken from different individuals
171. Exploratory work around GMDS applied solely to the nose region of different people (left and right), shown from different angles
172. Initial experiments with the Texas3DFR Database excluded the cheeks, which were later added as various parameters were studied for their impact
173. Texas3DFR Database pairs with the correct correspondence
174. Model modes with more than 1% variation built from correct pairs
175. Model modes with more than 1% variation built from false pairs
176. With GMDS issues still in tact, the ROC curve for recognition suffers
177. A pair that GMDS usually fails on
178. Another pair that GMDS usually fails on
179. A closer, GMDS-style look on the very flawed correspondence-finding (example from Figure )
180. Another example of a GMDS-type comparison applied to a real pair and failing
181. Pairs of facial expressions from the same person, cut in half beneath the nose and tilted sideways, then shown with GMDS applied. The score (from left to right): 3.4866, 2.4497, 2.4718, 10.9726, and 173.6779
182. Some of the raw images (full face) after GMDS with 50 Voronoi cells displayed
183. A top-down view showing the matching and the corresponding score. with flipping manually corrected. The third example from the top got the topology completely upside down.
184. 5 examples of experiments with synthetic data, where the top part shows the pair of images in their classic form, the middle shows a top-down view, and the bottom part is the range image
185. The scores in black show the pairings between different people and in green are the scores of matches between the same person
186. A new visualisation form where the dots signify stress at the given point
187. Top: A mapping of GMDS stress when cheeks are included in match-finding. Bottom: same as above but cheeks excluded. One can assume dark means low stress and white is high stress.
188. Original images, erroneous cropping effects (still in the process of debugging) and retriangulation of the points after omission.
189. A toy example of a very small couple of surfaces cropped from the centre of a face of the same person, where the pairs shown correspond to top-down view and GMDS' results
190. Example of an augmented slice from a pair of faces and GMDS applied to these
191. Results of a comparison between arbitrary bits where some boundaries are a Euclidean cutoff and some are geodesic
192. Results of a comparison between consistently chosen bits (near the eye) where some boundaries are a Euclidean cutoff and some are geodesic
193. Results of a comparison between surfaces that are mostly carved out of a geodesic boundary
194. Results of a comparison between noses with a boundary defined by geodesic distance constraints
195. A preliminary look at a predominantly geodesic mask and how it separates pairs from different people (top) and pairs from the same person (bottom)
196. An experimentation with a mask that includes points around the forehead and around the nose
197. With just 600 vertices, the ROC curve shows unimpressive ability to distinguish between true pairs and false pairs
198. The effect of increasing the number of vertices to 2420.
199. Improved performance with slight changes in surface size for the gallery
200. The result of changing the border threshold for surface carving
201. The result of growing the surface too big
202. Performance with (left) and without averaging (right) of the arrange image for better sampling of the GMDS process
203. A look at slicing at geodesic boundaries around the nose tip, with coarse resolution on the left and improved resolution that isolates regions on the right
204. Newer Fast Marching algorithm as applies to a face from the Texas database
205. Newer Fast Marching algorithm as applies to TOSCO dataset
206. A bug with connected triangles
207. Two faces and the issue with triangulation
208. The data (top) and the inherent bug which incorrectly connects points (bottom)
209. A correctly connected pair of faces with the source point highlighted
210. GMDS using the older FMM implementation superimposed on top of new and incompatible code
211. Visualisation of the increasing number of vertices
212. Coarse resolution performance compared
213. A finer resolution-oriented set of results obtained from fewer runs than before
214. An example where the hair entering the surfaces can interfere with GMDS-based recognition (GMDS as a similarity measure)
215. Example where hair is at risk as being treated like skin surface, depending on the mask/s
216. The result of the nose tip being misplaced (original on the left, after masking on the right)
217. The problem of non-overlapping faces, a result of misregistration/misalignment
218. Registered and correctly aligned image
219. Example of a correctly sliced image subset (before geodesic boundaries cutoff)
220. High-density (vertices) surface and the images it is carved off
221. Problematic image pairs
222. A pair that causes GMDS to fail
223. Problematic real pair (same person) where GMDS works but poorly so
224. A 3-D representation of a pair of images from the same person
225. GMDS failing to work as expected
226. A problematic pair which is seen as too different to quality as a match
227. ROC curve based on the smoothed surfaces variant of the algorithm
228. Example of some GMDS (mis)matches in the initial experiments
229. ROC curve based on the improved smoothed surfaces and somewhat better resolution
230. 3 problematic image pairs
231. The area of collision in GMDS-based face detection
232. Examples of shape pair residuals and the corresponding ROC curve
233. Residual difference and the problem of localised high signal (which makes this a weak similarity measure)
234. ROC curve obtained by using a residuals of just a particular image region (nose and eyes)
235. Examples of pixel differences for pairs of the same people
236. ROC curve corresponding to pixel differences for the whole middle section of the face
237. ROC curve corresponding to pixel differences for the nose area alone
238. ROC curve corresponding to sum of squared differences for the nose area alone
239. Example of 2 pairs from which the difference image is produced (shown at the top)
240. Top images show the sum-of-squared-differences of the first 3 true pairs, with the mere difference shown at the bottom
241. Examples of the first 12 false pairs (sum-of-squared-differences)
242. ROC curve generated by a sum-of-squared-differences-based similarity measure
243. ROC curve generated by a sum-of-squared-differences-based similarity measure
244. Examples of matches between true pairs and other matches between false pairs (different people). The separation is not yet profound enough to get state-of-the-art recognition performance.
245. Example of similarity values after a Euclidean delimiter (above the eyes) was removed
246. Example of similarity values with more points
247. Number of points pushed higher towards 350 (near the maximal allowed value)
248. Pairing examples with false pairs, 1000 vertices on each
249. Pairing examples with false pairs, 2000 vertices on each
250. At the top left is just a naive implementation, the top right shows what happens when GMDS failures get detected and removed, and the large plot shows what happens when Euclidean measures are factored into this toy example.
251. 4 problematic image pairs
252. Manually-measured width values for pairs of faces corresponding to different people
253. Manually-measured width values for pairs of faces corresponding to the same person
254. A geodesic ring/circle-based measurement as applied to tell apart anatomical equivalents from inequivalents
255. FMM is being used in a level sets-esque approach
256. An extension of the original (first) experiment which explored FMM (with Euclidean measures) as a classifier
257. Results from an extension of the range of radii/distances traversed from 20 to 50
258. Interim results (70 images) show 95% recognition rate with FMM-only (no GMDS) utility, but this tends to degrade as more difficult images are presented. Two good recognisers (classifiers), one of which is a Euclidean-geodesic hybrid, might give pretty good and mutually-independent results without using texture or fiducial points.
259. An example of two images from two different people, which nonetheless the FMM-based recogniser cannot quite detect as being different
260. An FMM-based recogniser results in nearly 90% recognition rate now (without GMDS)
261. The FMM recogniser ROC curve after increasing the number of true pairs
262. Example of a 10 degrees tilt
263. Recognition results from tilting one corresponding eye 360 degrees, then measuring distances on the geodesic boundaries
264. ROC curve based on comparing 220 images, where their Euclidean properties are measured upon geodetic slices
265. Brute force implementation that measures many geodesic distances
266. This figure visualises the idea of encoding surfaces as a vector not of surface vertices but an ordered list of Euclidean-upon-geodesic distances, which are fast to compute and sensitive to isometric/mildly detectable alterations
267. Separability testing in hyperspace
268. The image set of the first imaged individual in then test set, as an animation. The animation of the data from the 95th person is originally a GIF file.
269. Animation of the data from the 96th person
270. ROC curve obtained by measuring geodesic-Euclidean distances on the first imaged individual vs the same on different individuals
271. Smoothing versus no smoothing before measuring distances for identification purposes
272. The result of running the test set further (not for comparative purposes)
273. Animation of the data from the 103rd person. It is based on a set of images from the same person (numbered 103), without particularly challenging variation
274. The ROC curves based on a comparison between arbitrary (non-identical) pairs and pairs of images from the 103th person
275. The results following an increase in the smoothing range, demonstrating significantly degraded performance
276. An illustration by example of some images that prove to be challenging in the sense that their intrinsic properties are so similar that they almost get classified as being the same person (depending on how the threshold gets set)
277. Detection rate of my FMM-based method without GMDS as fallback (just annulling cases where fallback is invoked). X is log-scaled.
278. Results of GMDS applied to one single region rather than many, demonstrating the importance of having enough samples
279. FMM-based method without the use of ranges for fallback (and with some errors in pairs, which degrades the quality)
280. The result of applying the new method to pairs it feels confident enough comparing, based on pre-supplied thresholds
281. The problem with GMDS not finding a path through the graph in some cases, where eye regions get altogether cropped out as a result
282. The performance one gets by handing difficult cases based on nose alone or eyes alone. The results from GMDS are similar to the results attained using the other method which is still undergoing development and gradual improvements, maybe with exact geodesics.
283. The performance attained by removing hard cases
284. The performance attained by changing the number of vertices and keeping all pair examples to be judged for similarity
285. Example of poor alignment in the original set
286. Two examples of easy matches from the remainder of the dataset (which was enrolled in its entirety into the experiment)
287. Example of geodesic differences map around the nose (to be improved)
288. Example of a thin FMM spiral
289. Four small examples of distances spiral in isolation
290. 2 larger examples of distances derived from pairs of images of the same people
291. The distances spiral overlaid on the images it corresponds to (9th person in the set)
292. The distances spiral overlaid on the images it corresponds to (13th person in the set)
293. The distances spiral with larger, clearer points
294. Example of a true pair (same person) with a simplified representation of distances around each source point (FMM)
295. Another example of a true pair with simplified representation of distances around each source point (FMM)
296. An expanded view on the mis-correspondence between regions, where brighter shades represent greater disagreement between the pair taken from the same person
297. An expanded view on the mis-correspondence between regions, where the pair taken is from different people
298. Overview of the debugging process with examples from two true pairs (same person) and one false pair (different people), with the eye component discrepancies shown at the top and the nose at the bottom
299. The ROC curve obtained by using a weighted form of the similarity measure
300. ROC curve for the first phase of the experiment, which compares one-to-one (same person) and many-to-many (different people excluding this person, except in one case)
301. A broader scope curve for performance as in the previous figure
302. An example of misalignment in some parts of the nose in a true pair of images (same person), with the left nostril being a prime example
303. An example of a borderline case (leaning towards false positive)
304. Debugging information for the problematic true pair shown before
305. Debugging information (distance differences) for the aforementioned false positive
306. The problematic (borderline) false positive after the new alignment scheme gets applied
307. A contour around nose tip candidates all of which share the same (maximal) depth value, resulting in uncertainty
308. The Performance attained in hard cases where the tip is determined more arbitrarily than in a sophisticated fashion
309. The Performance attained in hard cases where the tip is chosen based on the average location of tip candidates
310. The result of applying a faster calculation of similarity, as applied to the first person against 90 different pairs from 90 different people
311. Performance when smoothing gets disabled, demonstrating little difference compared to prior results
312. The sort of results we get by using spectral masks without proper adjustment to make the masks shrewd enough. There is some potential there.
313. Example raw slice of the face of one subject
314. A test run with just one ring around the nose as a discriminant
315. The ROC curve obtained by using one single spectral/diffusion ring
316. The ROC curve obtained by using one single spectral/diffusion ring, applied only to the true positive gallery in the set
317. The ROC curve obtained by using just one diffusion distance as a discriminant
318. The ROC curve obtained by using two diffusion distances as discriminants
319. ROC curve for a rather disappointing approach of mask dilation based on diffusion distance
320. The same GUI in late March
321. Face cropping for standard experiments data
322. Difficulties identifying faces in GIP data
323. Difference images of the entire face surface before and after ICP-based registration
324. Assumed mark (left) extracted from accompanying GIP data (right), illustrating mis-detection
325. The offset problem visualised
326. The face before (left) and after cropping (right)

``The surest way to corrupt a youth is to instruct him to hold in higher esteem those who think alike than those who think differently.''
- Friedrich Nietzsche.