question

NserEric-3453 avatar image
0 Votes"
NserEric-3453 asked RohitRangarajan-7015 commented

Best practices to train model with many similar classes

Hello,
I am currently creating a model for the classification of clothing with Azure Custom Vision. In the process, I have some projects which need to distinguish many similar classes. For example, I have a project for classification of outerwear. This contains very similar classes like V-neck T-shirt, polo T-shirt, pocket T-shirt, etc.
What is a good approach to classify these very similar classes with one model as accurately as possible? Simply uploading more data doesn't help anymore, unfortunately.

I would be very happy to get an answer.
Best regards
Eric

azure-custom-vision
· 2
5 |1600 characters needed characters left characters exceeded

Up to 10 attachments (including images) can be used with a maximum of 3.0 MiB each and 30.0 MiB total.

@NserEric-3453 I am not sure if you have already tried this scenario but since this is a classification project, for a scenario like yours I would go for Retail domain with multilabel classification type where we can tag more than one label to an image.

For a T-shirt I would classify or label them based on the neck type, sleeve length, Fit, Pocket, Hood, design & Print.
The labels could be different variations of these types which can be assigned to images to train a small data set for the first iteration to perform a quick test. With only one negative tag that can be added, I would name the negative tag as Other and tag images which do not fit the list as Other because these are very similar datasets.

If you have already tried this before it would be great if you could provide more details of what did not work to advise any alternate ways to achieve the same.

2 Votes 2 ·

@romungi-MSFT How can we do multi-label classification using ML designer? I mean, how to represent the dataset? I see that the "Train Model" component does not allow the target label column to have a list of classes (when a text can belong to more than one class).

0 Votes 0 ·
NserEric-3453 avatar image
0 Votes"
NserEric-3453 answered

Hi @romungi-MSFT,
thanks for the recommendations. I've already set the domain to retail.
I have not yet tried multi-label classification. This is unfortunately not possible with my data, because I have no information about sleeve length, fit, pocket, hood, design & print. For other categories like shoes or pants, I also have to distinguish between many similar categories. There I have again other properties than sleeve length, fit, pocket, hood, design & print.

So far it has increased the accuracy if I have as few classes per project as possible. I also add blurred and grayscale images of the items to the projects to train the model with a higher variety of images. Furthermore I, classify the images in a hierarchical arrangement of many projects. For example, a t-shirt is classified as follows: Clothing --> Outerwear --> Shirts ---> T-shirts.

5 |1600 characters needed characters left characters exceeded

Up to 10 attachments (including images) can be used with a maximum of 3.0 MiB each and 30.0 MiB total.

KuehneKlatten-5338 avatar image
0 Votes"
KuehneKlatten-5338 answered RohitRangarajan-7015 commented

Thanks for sharing. This solution is really helpful for my project Lil peep apparel categories.


· 1
5 |1600 characters needed characters left characters exceeded

Up to 10 attachments (including images) can be used with a maximum of 3.0 MiB each and 30.0 MiB total.

@KuehneKlatten-5338 Did you implement multi label classification using Azure ML Designer?

0 Votes 0 ·