Q2. What is the main performance measure for their verification systems?
The main performance measure for their verification systems is the equal error rate (EER), which is the error rate at the decision threshold when the false rejection rate (FRR) is equal to the false acceptance rate (FAR).
Q3. How is the graph-based dissimilarity score obtained?
the graph-based dissimilarity score dGED is obtained by dividing the actual GED by the maximum GED, which is the cost of deleting all nodes and edge in the first graph and inserting all nodes and edges of the second graph.
Q4. What are the two alternative families of graph matching that differ from more traditional approaches?
Two alternative families of error-tolerant graph matching that differ in their basis from more traditional approaches, are graph embeddings and graph kernels.
Q5. What are the parameters used to determine the parameters for their graph-based method?
To determine the best parameters for their graph-based method, the authors performed a grid search with the following parameters: D ∈ {25, 50, 100}, Cnode ∈ {12.5, 25, 50, 100}, and Cedge ∈ {0, 12.5, 25, 50, 100}.
Q6. What is the simplest way to train a deep CNN?
The authors are considering two state-of-the-art CNN architectures:• ResNet-18 proposed by He et al. (2016), which is the 18 layer deep variant of a CNN that uses skip connections between layers to tackle the vanishing gradient problem.•
Q7. How is the biometric authentication system likely to improve?
the robustness of biometric authentication is likely to further improve when using a large multiple classifier system that combines even more structural and statistical classifiers.
Q8. What are the two types of forgeries used in the pattern recognition community?
The authors are considering two types of forgeries, which are common in the pattern recognition community, skilled forgeries (SF) and so-called random forgeries3(RF).
Q9. How can the authors train a network to verify signatures?
As the authors do have images of signatures, the authors can formulate the signature verification task as an image matching problem and proceed to train their network with the triplet-based method.
Q10. Why did Maergner et al. choose DenseNet-121?
The reason behind this choice is the particular nature of the DenseNet architecture, which al-lows features from lower layers to be propagated directly to the higher layers of the network.