Automatic face recognition for film character retrieval in feature-length films
read more
Citations
Probability and Random Processes
"Hello! My name is... Buffy" - Automatic Naming of Characters in TV Video
Face recognition based on image sets
2D Articulated Human Pose Estimation and Retrieval in (Almost) Unconstrained Still Images
References
The Nature of Statistical Learning Theory
A Tutorial on Support Vector Machines for Pattern Recognition
Robust Real-Time Face Detection
Introduction to Modern Information Retrieval
Robust real-time face detection
Related Papers (5)
Rapid object detection using a boosted cascade of simple features
Frequently Asked Questions (15)
Q2. What are the future works in "Automatic face recognition for film character retrieval in feature-length films" ?
The main research direction the authors intend to pursue in the future is the development of a flexible model for learning person-specific manifolds, for example due to facial expression changes. The authors are very grateful to Mark Everingham for a number of helpful discussions and suggestions, and Krystian Mikolajczyk and Cordelia Schmid of INRIA Grenoble who supplied face detection code.
Q3. What is the effect of the facial features on the image plane?
Seeing that the surface of the face is smooth and roughly fronto-parallel, its 3D motion produces locally affine-like effects in the image plane.
Q4. What is the main research direction the authors intend to pursue in the future?
The main research direction the authors intend to pursue in the future is the development of a flexible model for learning person-specific manifolds, for example due to facial expression changes.
Q5. What is the way to correct for the effects of varying pose?
After the face detection stage, faces are only roughly localized and aligned – more sophisticated registration methods are needed to correct for the effects of varying pose.
Q6. How was the effect on between-class distances found to be statistically insignificant?
The proposed approach achieved a reduction of 33% in the expected within-class signature image distance, while the effect on between-class distances was found to be statistically insignificant.
Q7. What is the last step in processing a face image to produce its signature?
MF = M ∗ exp− (r(x, y) 4)2 (8)IF (x, y) = IR(x, y)MF (x, y) (9)The last step in processing a face image to produce its signature is the removal of illumination effects.
Q8. What is the expected score of a random ordering?
The score of ρ = 1.0 can be seen to correspond to orderings which correctly cluster all the data (all the in-class faces are recalled first), 0.0 to those that invert the classes (the in-class faces are recalled last), while 0.5 is the expected score of a random ordering.
Q9. How many facial features are used in the film?
For training the authors use manually localized facial features in a set of 300 randomly chosen faces from the feature-length film “Groundhog day”.
Q10. How many people are in a typical film?
In a typical feature-length film the authors obtain 2000-5000 face images which result from a cast of 10-20 primary and secondary characters.
Q11. How do the authors suppress the image information around the boundary?
Foreground/background segmentation produces a binary mask image M. As well as masking the corresponding face image IR (see Figure 10), the authors smoothly suppress image information around the boundary to achieve robustness to small errors in its localization:
Q12. What is the method for removing particular distortions in the image?
The proposed approach of systematically removing particular imaging distortions – pose, background clutter, il-Proceedings of the 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’05) 1063-6919/05 $20.00 © 2005 IEEElumination and partial occlusion has been demonstrated to consistently achieve high recall and precision rates.
Q13. What is the way to train a facial feature?
In uncontrolled imaging conditions, the appearance of facial features exhibits a lot of variation, requiring an appropriately large training corpus.
Q14. What is the way to detect the face outline?
In detecting the face outline, the authors only consider points confined to a discrete mesh corresponding to angles equally spaced at ∆α and radii at ∆r, see Figure 9 (a).
Q15. What is the effect of the affine transformations on the face image?
Noting that these produce mostly slowly varying, low spatial frequency variations [11], the authors normalize for their effects by band-pass filtering, see Figure 3:S = IF ∗ Gσ=0.5 − IF ∗ Gσ=8 (10) This defines the signature image S.In Sections 2.1–2.4 a cascade of transformations applied to face images was described, producing a signature image insensitive to illumination, pose and background clutter.