Inference of Drawing Elements and Space Usage on Architectural Drawings Using Semantic Segmentation
Round 1
Reviewer 1 Report
Dear Authors,
Line 90 – The paragraph should start with a capital letter
Line 61 - Acronyms such as CNN, FCN, VGG, GAN, etc., must be clarified the first time they are related in the text or add a glossary.
Line 228 – “EXPERIMENT” Adapt the section title to the instructions.
Line 250 – “Figure 4. Conversion Proces.” instead of “Figure 4. Conversion Process.”
Line 295 – “In the encoder, it is rst bilinearly ..” instead of “In the encoder, it is first bilinearly..”.
Line 439-441 - Adapt the reference number 22 to the instructions for authors.
I think that is a good research paper, with an appropriate research design, the model structure and experiments adequately described, and the results and conclusions are clearly presented.
Author Response
Dear Reviewer,
We would like to sincerely thank you and the editors of the Applied Science for taking the time to review our article. We have made some corrections and clarifications in the manuscript after going over the your comments, and we believe that these have greatly improve our revised manuscript. Please find our point-by-point responses to each of the your comments below:
#[RE: Comment 1] (Line91) Has been modified.
#[RE: Comment 2] (Line61,62,180,166) CNN, FCN, etc. are terms that are used a lot in deep learning, so we didn't think about it. We agree with your opinion. Each of the abbreviations were added.
#[RE: Comment 3] (Line230) Has been modified.
#[RE: Comment 4] (Line252) Has been modified.
#[RE: Comment 5] (Line302) Has been modified.
#[RE: Comment 6] (Line446) Has been modified.
We are very grateful for your thorough and friendly review and would like to thank you for taking the time to review our article. Your comments have provided us with a good opportunity to improve our research even further going forward.
We hope the revised manuscript will better meet the requirements of your journal for publication. We thank the editor and the reviewers of the Applied Science once again for the constructive review of our paper.
Sincerely yours,
Jihyo Seo on behalf of the authors.
Corresponding author: Seungyeon Choo at School of Architecture, Kyungpook National University, 80, Daehak-ro, Buk-gu, Daegu, 41566, Republic of Korea, [email protected]
Reviewer 2 Report
The paper proposes an automated method based on deep learning with the aim of automating the architectural design. The techniques are classigiciation, localization/detection and segmentation are used. Introduction and related work are widely explained. However, it is not clear the contribution of the present work with respect to the current state of the art. As well, how the work done benefits from the previous researches. A "beyond state of the art" is encouraged.
Moreover, it is also encouraged the explanation of the methodology that has been applied.
Some other comments are:
- Semantic segmentation: Is there any standard data model that has been used for the representation of the information, e.g. BIM and/or relating the BIM data with the objects' "pixel"? Apart from the JSON way to publish the information, does it follow any standard data model like IFC or is it just JSON information?
- How is the detected information from the deep learning model mapped into building information? That is to say, when detecting a door, how is this door classified?
- This is key for automated design as the architecture design is not just the geometrical or elements, but also their feature and its link with BIM families. Is this covered?
Author Response
Dear Reviewer,
We would like to sincerely thank you and the editors of the Applied Science for taking the time to review our article. We have made some corrections and clarifications in the manuscript after going over the your comments, and we believe that these have greatly improve our revised manuscript. Please find our point-by-point responses to each of the your comments below:
#[RE: Comment 1~2 (the contribution of the present work and the explanation of the methodology)] In this study, we built a network that infers drawing elements and usage of space through ‘transfer learning’ that re-learns by changing a specific layer to an existing learned model, rather than creating a new layer from scratch. Transfer learning can be implemented quickly and easily by using a pretrained network as a starting point. By training the network on a new data set, it has the advantage of fine-tuning the deeper layers of the network, while at the same time creating a network specific to your own new data set. Transfer learning is already being used in many studies, and meaningful results are being produced through it. We agree with your opinion that an explanation for this should be added, and we have written the methodology more specifically. (Line295~300)
[RE: Comment 3~4 (standard data model and information mapping)] This experiment works on an image basis. Training is taking place by recognizing the shape of an image as a pixel, and it does not have information like BIM data. It recognizes the shape based on an image, not an object-oriented data model such as an IFC file, and classifies it into a defined class. Defined classes are the elements we want to classify. In the case of drawing elements, there are 5 classes: wall, window, door, sliding door, and emergency evacuation door, and for space use, there are 8 types of room, entrance, balcony, dressing room, bathroom, living room, evacuation space, and pantry. This is a class that we have defined in advance, it only learns and classifies images, but the concept is slightly different from the object-oriented model of IFC. The json file is a Javascript-based file format made to be convenient for sending and receiving a large number of files as described in Section 4.1, and the json file used in this study is a large number of image files.
#[RE: Comment 5 (linkage with BIM families)] We fully agree with your opinion. It is difficult to say that it is an automated design only by automatic recognition of the floor plan. Of course, as you said, It also need to connect with the BIM product line. Accordingly, We know that research on automatic recognition of IFC objects is also being actively conducted. As a follow-up study of this study, we are conducting research to allow artificial intelligence learning the floor plan to automatically generate spatial plans, and after that, research should be conducted to design and implement automation through connection with BIM products. However, the method proposed in this study is the basic stage of design automation, and helps to build a large amount of learning data sets by learning many architectural drawings and automatically obtaining labeled data. We think this is meaningful in that it can lay the foundation for architectural design automation.
We are very grateful for your thorough and friendly review and would like to thank you for taking the time to review our article. Your comments have provided us with a good opportunity to improve our research even further going forward.
We hope the revised manuscript will better meet the requirements of your journal for publication. We thank the editor and the reviewers of the Applied Science once again for the constructive review of our paper.
Sincerely yours,
Jihyo Seo on behalf of the authors.
Corresponding author: Seungyeon Choo at School of Architecture, Kyungpook National University, 80, Daehak-ro, Buk-gu, Daegu, 41566, Republic of Korea, [email protected]
