Q1: Why a finer threshold (0.99) could not find one of the 2-humped camel?

Ans: The higher threshold means the higher similarity it should have between the template and the searched image. At threshold of 0.99, that camel can't match because it may have some missing pixel.

Q2: Is it possible to have decimal points for threshold?

Ans: Yes, you can try to use different decimal points and observe the changes.

Q3: Could you explain what will happen if we change minNeighbors?

What about "minNeighbors" parameters? How should we change it?

Ans: 

scaleFactor: Parameter specifying how much the image size is reduced at each image scale. Increasing it leads to faster detection with the risk of missing some objects, whereas a small value might sometimes be too thorough.

minNeighbors: Parameter specifying how many neighbors each candidate rectangle should have to retain it. Higher value results in less detection but with higher quality.

minSize: Minimum possible object size. Objects smaller than that are ignored.

maxSize: Maximum possible object size. Objects larger than that are ignored.

Q4: Why  scaleFactor 1.1 and 1.9 is same result? there are no change result

Ans: It's because the rectanglea has combined in original image. If you want to get exact result you have to back to the previous step and run it. Then, you can run the paragraph of  new scaleFactor and get the right result.

Q5: I tried to change minNeighbors from 5 to 2 and it detected other objects apart from the faces. Could you please explain that effect?

Ans: minNeighbors: Parameter specifying how many neighbors each candidate rectangle should have to retain it. Higher value results in less detection but with higher quality.

Q6: In real life, how many epoch will be train or when will be stop training?

Ans: Generally, when the learning curve goes on an incrementally smaller scale and reaches a plateau, the epoch value in the moment is the appropriate one.

Q7: how much data should we put in training data set/testing data set?

Ans: Depending on your dataset size and use case, as shown in the next answer.

Q8: What is the general rule of thumbs for picking the right numbers of samples, for separating the train and the validate samples?

Ans: There are three cross-validation methods that we usually use in deep learning: 1) k-fold cross-validation, 2) leave-one-subject-out cross-validation, and 3) leave-one-out cross-validation.

k-fold cross-validation: Most commonly used in machine learning and deep learning. In k-fold cross-validation, the original sample is randomly partitioned into k equal sized subsamples. Of the k subsamples, a single subsample is retained as the validation data for testing the model, and the remaining k − 1 subsamples are used as training data. k is determined by users, usually set k as 5, 7, 9, 10.  

leave-one-subject-out cross-validation: Similar to k-fold cross-validation, each subset contains data from the same subject. You will have the same k as the subject number of the dataset.

leave-one-out cross-validation: Keep only one image as a testing set, usually used in a small dataset (less than 1000 images).

REF: TTAIC & AIGO & III