The weights in W can be negative.
The comparison to other works on different databases is still problematic.
How would the method behave with kernels of different size?
I don’t understand why you use the negative gradient.
Is it for humans?
I would appreciate if you could create a concise table.
My impression after reading the article is that you are going to make the dataset public.
The article seems to propose replacing the method of approximate nearest neighbor search. However, the article is written poorly and incomprehensibly.
Does it make sense to have negative weights?
Equation 5 does not include translation which the text suggests it does.
The article should be largely rewritten before any resubmission.
Its observation probabilities are computed as a product of pixel probabilities.
Anyway, what is the motivation for equation 13?
It would help the reader to understand your method better.
The data from Figures 5 and 6 should be represented much better.
I would argue that much better global and local contrast enhancement methods exist.
Image registration could be easily incorporated into low-power embedded devices.
Without the dataset, the article is useless.
Every image processing software contains automatic contrast and color enhancement.
Why are the weights computed this way?
The article references previous biometric work in a reasonable way.
The review of related methods is simple to understand and informative.
No one can verify your results.
Most algorithms are not very sensitive to the type of normalization used.
The article should discuss possible approaches to image registration.
It considers the horizontal dimension as time.
It could be computed in many ways which would give more or less equivalent results.
He does not need to understand how the transformation is computed.
The text is often unnecessarily obscure.
You should show that your method improves results.
The text does not discuss the meaning and effect of the weights.
The references are good in the very narrow area of histogram methods.
I understand that it is your dataset, and that it is new.
The paper presents a variant of a hidden Markov model for images.
The article addresses image classification with local image descriptors.
I believe that this is not explained in the text.
No one can compare to your results.
I don’t know how comfortable it would be for a person.
Is it for further automatic processing?
I can not help but feel that this topic is hopelessly outdated-
The experiments are unconvincing.
The only meaningful results are currently in Table 9.
It should provided results comparable to the state of the art.
The probabilities are computed for larger image patches and using quantized representation of the image patches.
I feel quite ambivalent about the article.
The text implies that method handles scaling, rotation, and translation.
The fact that you can write equations the way you do does not mean you should to it.
The experimental methodology lacks separate validation set.
Explanations are illogical and confusing.
You should evaluate it in a perceptual experiment.
It is true that some methods benefit from normalization of image inputs.
No baselines or comparison to state of the art is provided.
However, the authors ignore a very large body of work on tone mapping.
Could you provide an example of a realistic scenario?
You don’t compare directly to existing state-of-the-art methods.
The ideas should be presented in a more concise and direct manner.
Rarely have I seen such a collection of poorly motivated simple methods.
The model has effectively three fully connected layers.
I failed to reach a conclusive interpretation.
The presented method can be applied to more realistic scenarios with a fixed camera.
I have a problem with how you motivate the method.
The achieved results are far behind the state-of-the-art in image classification.
I still find the article to be unnecessarily obscure.
You have missed more than two decades of rapid progress in image processing.
Your method is systematic and mostly reasonably motivated.
However, the presentation of the previous work could be deeper.
The model can be initialized with sparse coding.
Without the dataset, the article would be meaningless.
I found the compression phase problematic.
Doesn’t it defeat the purpose of hashing?
The discrepancy between reader’s expectations created by the text and the real behaviour should be resolved.
The article in the current state is not good enough for publication.
I find the other argument dubious.
Evaluation should be extended to other datasets.
Formatting of tables should be improved.
The article would benefit from examples demonstrating the dataset weighting.
The article should contain URL where the dataset can be downloaded.
How are the weights computed?
It could have been written much more clearly.
Isn’t this a parameter of the method?
Some of the presented conclusions are not supported by the reported results.
This sentence does not make any sense.
What is the benefit of the method in terms of identification accuracy?
The parameters are selected on the test set.
I would not recommend publishing the article without the dataset.
I find the need for biometric identification from palm images taken by infrared handheld camera questionable.
The article introduces a very fast, small, and shallow feed-forward network which is trained end-to-end.
The proposed approach provides state-of-the-art results at very low computational cost.
The authors claim to specifically address variations due to geometric transformations in the acquired images.
The reader just needs to know what it does.
The full network is trained to minimize L2 pixel loss.
The usage of high information content is meaningless.
The reader does not need to know how it is actually computed.
The pixel probabilities are conditioned on the model state.
The method does not have to be the best.
I tried to understand the proposed application of to the best of my ability.
Image registration and invariant representations could both be done in a computationally very efficient way.
How do you handle the negative weights?