Trying to classify a rather messy image dataset

vision
1

Hi,

This post is not specifically about a pytorch related issue but a general computer vision problem and I hope this is the correct place for such a discussion.

I am trying to classify a rather messy dataset of images. The images are taken using mobile phones and scanner scans from various sources so there is some variation in the dataset.

Its a dataset of election party flyers/leaflets, there are about 5 classes and each class/label has about 3000 images associated with it so there is enough variation here. Here are a couple examples of what the images look like:

leaflet_1041_1
leaflet_1041_11024Ă—1400 223 KB

leaflet_100_2
leaflet_100_21449Ă—1024 204 KB

leaflet_1139_2
leaflet_1139_21024Ă—1437 183 KB

leaflet_1015_1
leaflet_1015_11024Ă—1096 185 KB

As a starting point, I used a ResNet 101 network to learn this dataset and for preprocessing all I did was resize the images to 200 x 200 using interpolation.

But after training for over 24 hours the validation/test accuracy is about 54%, so the network did not lean much and is just random guessing.

Is it possible to classify such a dataset? Any tips on how to pre-process these images to improve result?

Thanks

2

That doesn’t seem to be the case for 5 classes, as a random guess would give you an accuracy of 20%, wouldn’t it?

I think there might even be “too much variation” given the example images and the comparably small dataset size.
However, I’m quite impressed that the model has learned something at all (if I’m not mistaken in my first point).

1 Like