Used clothes - image classification

Hi Everyone,

I have images from a dataset which consists of used clothes looking like below:

432675370![581147844|198x264]label:3/4 sleeves

581147844 label: full sleeves

There are 4 classes (sleeveless, half sleeves, 3/4 sleeves and full sleeves). Using a pretrained resnet34 and the methodology described by @jeremy in the earlier lessons of part 1, I am unable to get a decent model, mainly due to the following issues (I think):

  1. The confusion matrix shows that the model is confused between 3/4 sleeves and full sleeves. And, as we can see from the images, these two classes are hard to discriminate.

  2. There are only 500 training images per class.

Does anyone have any inputs on how to deal with such a dataset?

I’d be confused too! :slight_smile:

If the classes are all about the sleeves, I’d try training a detector to find the bounding box (lesson 8) of either the clothing article or the sleeves themselves and cropping the images to regularise their proportions before feeding them to the classifier.

Another thought would be to look at the confidence level it’s is incorrectly giving full sleeves and see if you can clip at some level (eg 75%, 3/4!) and call them ¾.

Also resnet-34 may not be the best fit, try a densenet?

@digitalspecialists, thank you for your input. I will definitely try object detection on the sleeves. Any particular reason why you say that densenet would work better than resnet?