Cross-modal localization through mutual information
read more
Citations
Arbitrary Talking Face Generation via Attentional Audio-Visual Coherence Learning
High-Resolution Talking Face Generation via Mutual Information Approximation.
A dependence maximization approach towards street map-based localization
Dependence maximization localization: a novel approach to 2D street-map-based robot localization
References
Mutual information function assesses autonomic information flow of heart rate dynamics at different time scales
Current and future trends in sensor networks: a survey
From error probability to information theoretic (multi-modal) signal processing
Forest sampling desk reference
Entropy manipulation of arbitrary nonlinear mappings
Related Papers (5)
Frequently Asked Questions (15)
Q2. What are the future works in "Cross-modal localization through mutual information" ?
Research in several directions to extend the work presented in this paper are currently under way. Combining the indirect estimation methods with direct estimation could couple their respective strengths and would be a fruitful avenue of further research into signal grouping. Constructing a multidimensional feature space by combining the separate features could add value and this would obviously benefit future research outcomes.
Q3. How many pixels were used to analyze the video data?
Color images acquired were transformed to grey scale and pixel intensity values (consisting of 640 ∗ 480 = 307200 pixels per frame) of 100 frames were analyzed using raw pixel values.
Q4. What is the effect of a feature-level approach?
Formulating the problem in the feature level rather than signal level will remove the requirement of preserving locality of thedata source.
Q5. How many iterations did the L1 norm penalty produce?
Applying the L1 norm penalty to the optimization produced faster convergence, occurring in iteration 72 compared to 110 iteration with L2 norm penalty.
Q6. How many iterations of the optimization should be used?
In order to detect that the optimization has reached a local minima the variation of δ should be contained in a 1.5e−3 limit at least for a minimal convergence span of 5 iterations.
Q7. How can the authors achieve the maximisation of MI?
The maximisation of MI is achieved by maximising the entropies H(Y1) and H(Y2) and minimising the joint entropy, H(Y1, Y2) in (1).
Q8. Why do the other mapping parameters have smaller non-zero vales?
due to the approximations in the objective function and the presence of local minima, the other mapping parameters have smaller non-zero vales.
Q9. How can the authors maximize the entropy of the measure?
The entropies H(Y1) and H(Y2) can be maximised by selecting the mapping parameters to make the data on the lower dimensional space resemble a uniform distribution.
Q10. What is the criterion for imposing a penalty on the input space?
the solution of the parameter vectors α1 and α2 should be sparse identifying the minimum number of nonzero elements naturally suggesting the use of the L1 norm as an appropriate penalty function.
Q11. What is the simplest way to calculate the L1 penalty?
Since the projections α1, α2 may be of very high dimensionality, it is assumed thatmin ‖α1‖1 = |α11 | + |α12 | + · · · |α1n | (16)Therefore the L1 penalty is∂ min ‖α1‖1 ∂Y1(17)further∂|α1| ∂Y11 = ∑n i=1 ∂|α1i | ∂Y11 = ∑ |X−11 |row1 sign|Y11 | ... ∂|α1| ∂Y1i = ∑n i=1 ∂|α1i | ∂Y1i = ∑ |X−11 |rowi sign|Y1i |resulting in∂ min ‖α1‖1 ∂Y1 = ∑ |X−11 | sign|Y1| (18)All iterative optimization methods require stopping criteria to indicate the successful completion of the process.
Q12. How can the authors minimise the entropy of the measure?
joint entropy H(Y1, Y2) can be minimised by selecting the mapping parameters to reflect the joint distribution, (Y1, Y2) is furthest away from a uniform distribution.
Q13. Where is the laser beam of the range finder intersecting?
The laser beam of the range finder intersects horizontally at the abdominal area of the standing person capturing the movement of the book.
Q14. How can the authors get the entropy of the measure with respect to the mapping parameters?
They proposed an unsupervised learning method by which the mappings g1(·) and g2(·) can be estimated indirectly, without computing mutual information.
Q15. What is the difference between the two norms?
The L1 norm performs equally well as the L2 norm on overdetermined system of equations while outperforming L2 norm for underdetermined problems [9] especially where the solution is expected to have fewer non zeros than 1/8 of the number of equations.