Coupled information-theoretic encoding for face photo-sketch recognition
read more
Citations
DualGAN: Unsupervised Dual Learning for Image-to-Image Translation
Face Alignment at 3000 FPS via Regressing Local Binary Features
Multi-View Discriminant Analysis
Encoding in Style: a StyleGAN Encoder for Image-to-Image Translation
A Comprehensive Survey to Face Hallucination
References
Distinctive Image Features from Scale-Invariant Keypoints
Elements of information theory
Eigenfaces vs. Fisherfaces: recognition using class specific linear projection
Video Google: a text retrieval approach to object matching in videos
Relations Between Two Sets of Variates
Related Papers (5)
Heterogeneous Face Recognition Using Kernel Prototype Similarities
Frequently Asked Questions (13)
Q2. What future works have the authors mentioned in the paper "Coupled information-theoretic encoding for face photo-sketch recognition" ?
In the future work, the authors would like to further investigate the system with more cross-modality recognition problems.
Q3. What is the method for comparing pseudo photos to gallery photos?
Pseudo photos are synthesized from query sketches, and random sampling LDA (RS-LDA) [30] is used to match them to gallery photos.
Q4. What is the way to learn the projections?
At each node, the authors randomly sample α percent (empirically α = 80) of the element indices of the sampled vectors, i.e. use a sub-vector of each sampled vector, to learn the projections.
Q5. How do the authors train a coupled projection tree?
To train a coupled projection tree, a set of vector pairs X = {(xpi ,x s i ), i = 1, ..., N} is prepared, where x p i ,x s i ∈ RD.
Q6. Why do the authors choose 256 leaf nodes?
Due to small performance gain and high computational cost of a large leaf node number, the authors choose 256 leaf nodes as their default setting.
Q7. What is the main purpose of the first family of approaches?
The first family of approaches [27, 18, 31] fo-cused on the preprocessing stage and synthesized a pseudophoto from the query sketch or pseudo-sketches from the gallery photos to transform inter-modality face recognition into intra-modality face recognition.
Q8. What is the problem with the kernel trick?
kernel trick causes high computational and memory cost due to the very large size of the training set, and the nearest centroid assignment may be unstable (there is no hard constraint to require a pair of vectors in the same cluster).
Q9. How did Lin and Tang map features from two modalities into a common discriminative space?
In order to reduce the inter-modality gap at the classification stage, Lin and Tang [17] mapped features from two modalities into a common discriminative space.
Q10. What is the vote of a pixel to the histogram?
The vote of a pixel to the histogram is weighted by its gradient magnitude and a Gaussian window with parameter σ centered at the center of the region.
Q11. What global approach was proposed by Gao et al.?
Another global approach proposed by Gao et al. [9] was based on the embedded hidden Markov model and the selective ensemble strategy.
Q12. What is the ROC curve for the CITP forest?
The ROC curves are shown in Fig. 5. Even 32- dimensional CITE2 (please refer to Section 3 for this notation) significantly outperforms the 59-dimensional LBP and 128-dimensional SIFT.
Q13. What is the effect of different parameters on the performance of a CITP forest?
The authors investigate the effect of various free parameters on the performance of the system, including the number of leaf nodes, the projected dimension by PCA+LDA, the size of randomized forest and the effect of using different sampling patterns.