1. Feature based vs. holistic processing
Tanaka & Simonyi (2016), Sinha et al. (2006)
composite face effect
face inversion effect
whole-part effect

2. Feature based vs. holistic processing
composite face effect
face inversion effect
identical top halves seen as different when aligned with different bottom halves
when misaligned, top halves perceived as identical
whole-part effect
- methods talked about weren’t inspired by what we know about human ability to recognize faces, but raise some interesting questions
- two recognition methods - Kanade based on detailed analysis of parts of face, geometric arrangement, metric distances, angles between salient points on face, eigenfaces approach global method taking into account overall pattern brightness variations in face
- question asked in many perceptual studies, not fully resolved, whether recognition faces in human system based on feature vs. holistic process that takes into account more global, configural information about faces - some perceptual phenomena used to argue face processing human system more holistic, fuzzy notion in literature, show examples of effects that people point to
- composite face effect, look at left pair of images, tell me whether two top halves are same or different? identical tops, but when different bottoms aligned, can appear different (illusion, really same), when bottoms misaligned, more likely to make correct judgment, argued we derive more holistic representation of face when parts aligned that prevents access to detailed parts needed for comparison, demonstration with famous people, small information e.g. eye, sufficient to recognize famous person, mix woody allen/oprah winfrey disrupts recognition, interferes diagnostic capability of individual features
- inversion effect, more difficult to recognize inverted faces, disparity between our ability to recognize upright vs. inverted faces greater than disparity in recognition performance measured for other kinds of objects, argued inversion disrupts holistic arrangement of collection of features, eyes, nose, mouth, inversion effect develops early childhood, prosopagnosics don’t exhibit (learning process over time does not occur, contributes to difficulty)
- two weeks, see how model Tommy’s group proposed for neural computations underlying face processing can account for phenomena like this
inversion disrupts recognition of faces more than other objects
prosopagnosics do not show effect
Identification of “studied” face is significantly better in whole vs. part condition

3. The power of averages, Burton et al. (2005)
PCA performance
average “shape”
- Mike Burton/colleagues, internal representation of individuals based on average of examples seen, average images of famous people, how created, grabbed examples internet, roughly frontal views, otherwise unconstrained (lighting, facial expressions, time), not just average raw images, average texture (captures typical variation in brightness of face), average shape (shape manual step, generic grid face, move key points to corresponding location face), then (1) morph image to common geometric shape, do same for each exemplar, then average (individuals not all look like Tony Blair, average captured essence), but lost individual shape (large forehead), (2) average locations of key points on grid to get average shape captured by mesh, (3) morph average texture to grid  average face
- quite effective for recognition of famous faces – train PCA based recognition system with averages, performs better recognizing new individual images, improves with more exemplars combined in average
- interesting test with commercial system FaceVACS, database about ~ 31,000 photos of famous faces, ~ 3,600 celebrities, online version can upload image and returns closest matching photo from database (black box) – 460 individual images of famous people 54% correct, with average faces as input, reaches 100% correct, then computed average using 46% wrong/80% correct
- why? what doing with average that’s advantageous? averaging factors e.g. lightness variations, expression, not diagnostic of identity
- human vision, perceptual experiments on recognition of famous faces, average reaction times faster, less error ID average faces vs. individual exemplars, questions about how to automate construction of representation, how to evolve over time as become more familiar
average “texture”

4. Human recognition of average faces
Burton et al. (2005)
Performance: texture + shape images
Performance: shape-free images

5. Role of facial motion, Knappmeyer et al. 2003
animating faces
- study related to issue relevant when return to modeling recognition process in more biologically relevant way, prime with hypothesis explored in behavioral experiment Wallis & Bulthoff, some perceptual studies used different viewpoints, static views, most machine systems don’t use dynamic information from videos, just snapshots, presumably different views of same individual’s face are linked together, does seeing views in temporal succession facilitate linking?
- in experiment, ask whether can use temporal association to get people to incorrectly link two different faces, in way that disrupts ability to recognize whether two different views of same person (Tommy frontal slowly transform in motion, into Shimon’s profile, would that mess up my ability to later look at static snapshots of Tommy frontal and Shimon profile, and say, two different people)
- nice laser scanner can use on people, high accuracy 3D model, also scanned visual texture of face, 12 female faces, cropped hair so just face (CG look, but real images), scanned image texture from particular view, but given texture and 3D model, use graphics software to generate appearance of other poses, used morphing software to create sequences of 5 images, front one person, profile different person, 45 deg morph between two (3 rows, 3 groups, grouped differently for different subjects, faces within rows used to create morph sequences), training phase learn to recognize new faces, twice back and forth through sequence 5 images, then test, two static views one after other in time, same person? could be pair within same group that went through morphing, or pair across groups, never temporal association, reported data for mismatches, what predict results if hypothesis true? not correctly identified as mismatch (different, but if seen morphed, temporal association, think same person)

6. Role of facial motion, Knappmeyer et al. 2003