kumoh national institute of technology
Networked Systems Lab.

Ali Aouto, Thien Huynh-The and Dong-Seong Kim "Gait Features Extraction Technique for Identity Detection Using CNN in Military Surveillance Systems"
By : Ali
Date : 2020-03-25
Views : 184

Comments from the editors and reviewers:
-Reviewer 1



This paper proposes an identification system, which can train a sequence of gaits to recognize a person using a pre-trained CNN model. Especially, in order to recognize the unique sequence of gaits which a person has, the model of the proposed system leverages the orientation and distance of joints extracted from 3D skeleton information. Based on this gait-identification model, this paper evaluates the system on four different pre-trained models and 150 sequences collected from 30 participants (UPCV) dataset.


Overall, this paper introduces the gait-based identification system to overcome existing identification methods based on appearance features such as the face, fingerprint, etc. Through several references related to the existing works, this approach is enough to assert the necessity of the proposed system. But, I think that the other parts in this paper leave much to be desired. For this reason, this shortcoming also makes the contributions of this paper unclear. So, I'm worried about the following points.


-This section is now trying to deliver too much information including introduction, motivation, and related work. These parts can be shorten to address system design and contribution clearly.


-The mathematical symbols in the proposed scheme are not predefined. This makes the flow of mathematical equations hard to understand.

--- Why does the equation, 'h(h-1)', represent the amount of distance and orientation values? And please describe the detail information about 'h'.

-The proposed model is based on a pre-trained deep learning model. However, a series of these processes related to getting illustrative statistics characteristics are designed to make the dataset as an input for the deep model. These processes seem to make the deep learning model meaningless.

-There is no reasonable explanation of why the model pre-trained by image dataset was used. Also, rather than writing unnecessary equations, I recommend the authors need to address how to modify the model.


-Data collected from subjects is usually cross-validated by splitting subjects rather than by splitting data. I think the authors should include an additional description of this part.

-In order to recognize a particular person, it seems that this system must have that person's gait information. Unlike what this paper claims, how is it possible to identify strangers who are not included in the training dataset?

- Need to reorganize the list of the equations including a symbol table
- Need to define what are contributions to this paper
- Lack of details and discussions
+ Motivation and Not bad performance

-Reviewer 2

This paper proposes a gait-based user identification system. More specifically, the authors introduce a posture-based approach that uses geometric attributes (e.g., distances and orientations between joints) collected when users are walking. As a tool for identifying users based on the attributes, pre-trained CNN models are slightly modified for suiting the purpose of this work. The experiment result with the UPCV dataset shows a high degree of accuracy.

- The motivation is strong.
- The technical design, especially feature extraction, was described in detail, which was very informative.
evaluated and reported in detail.

1. The authors use pre-defined CNN models with slight modifications. But, details for this transfer learning process are not explained well. For example, in the transfer learning, we need to decide which parts (layers) will be re-trained with a new dataset. We can select some lower-level layers in a convolution base or only classification layers. It is unclear how the authors fine-tuned the pre-trained models.

2. The proposed system was evaluated with the UPCV gait dataset collected in controlled environments (during collecting data, users walk a straight line with a normal speed). However, in practice, there are several issues that cause a decrease in the quality of data (e.g., some joints are missed). I'm curious the performance of this system in realistic environments.

3. When reporting the evaluation results, could you compare the results with those from existing works (e.g., aspect-based approaches).

4. There are some major typos and grammar issues throughout the paper

Overall, I like the idea and design in this paper. However, there are some issues that prevents it to be published at this stage. Therefore, I encourage the authors to improve the paper and resubmit.