A Semi-Supervised Learning Approach for Automatic Detection and Fashion Product Category Prediction with Small Training Dataset Using FC-YOLOv4
Round 1
Reviewer 1 Report
The article introduces a semi-supervised pipeline for fashion detection. It built a dataset and a detecter based on YoloV4. Some issues should be addressed:
I am wondering whether only one target is labeled in each single image, even that there may be multiple object-of-interests in one image, e.g., pants and jacket. Additionally, it is not clear to me whether all object-of-interests are detected for each testing image. For examples in Fig. 13, only one result is provided for each image.
How will the different image augmentation techniques affect the performance?
The related work section is not comprehensive. Some human-centric analysis works should be included like Cascaded parsing of human-object interaction recognition and Differentiable multi-granularity human representation learning for instance-aware human semantic parsing.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
The authors Yamin Thwe, Nipat Jongsawat, Anucha Tungkasthan presented a a semi-supervised learning approach for automatic detection and fashion product category prediction with small training dataset using FC-YOLOv4. However, here are some minor observations:
Introduction is not exhaustive more details required with appropriate citations.
Also please separately add contribution of the study in Introduction section.
Mention the challenges faced while implementation and how did you tackle these
Mention if there are any trade-offs for implementation of the proposed approach while we deviate from traditional practices.
Some punctuation errors are observed. Please improve the punctuation in the manuscript.
Please mention experimental setup for performing the mentioned experiments and simulations.
Add some authors own insights about results in the Results section
Please elaborate conclusion by adding limitations of the proposed study.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
The paper addressed the problem of effectively fashion category detection for the field of business intelligence. To detect multiclass fashion products based on digital images, a model intuitively called FC-YOLOv4 (Fashion Category- YOLOv4), is proposed. In addition, the authors propose an image-based classification model utilizing FC-YOLOv4 to automatically recognize the product category for properly captured and clutter images. A semi-supervised learning approach is used to automatically annotating data, as well as image augmentation to increase the number of images in order to improve object detection accuracy. Results and discussions are provided to show how the FC-YOLOv4 model performs in comparison to the YOLOv4 and YOLOv3 category detection models.
The obtained results are particularly useful for second hand clothing or less professional marketplaces, where category information is difficult to be reliably available as opposed to vendor side properly photographed and category attributed data.
Bringing context and structure into focus is highly challenging future research direction of the authors.
Overall the paper is good written. All references are relevant and adequately used.
Remarks:
- in formula 1, Bn+1 could be misunderstood since Bn is the last bounding box of the collection, maybe using i and n like in Algorithm 1;
- line 416 – W^ and H^ are graphically not showing up correctly, not to be confused;
- algorithm 1, ENDIF should be line 20.
- figure 10(c) CLAHE results – is the visual representation in the correct colorspace for viewing;
- dataset is build upon publicly available images, since the authors make their results and datasets publicly available, it should be made clear that there are no legal concerns, if necessary.
Therefore, I consider that the paper “A Semi-Supervised Learning Approach for Automatic Detection and Fashion Product Category Prediction with Small Training Dataset using FC-YOLOv4” should be accepted after minor revision.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The revision has addressed most of my concerns. In the final version, please carefully re-check the reference list, in which some papers don't contain conference/journal information, e.g., [27][26][29][48].