Image classification Problem with > 1000 Classes

Part 1 (2019)
1

I try to identify 3000 tabletop games by images. I have an ImageBunch containing 3000 games and around 20 images per game.
Using Resnet50 (with only 100 different games) I can’t get an error rate lower then 60%. I tried different transformations like flipping, rotation, lightning, … without success.

Are cnn’s with resnetXX a good pick to identify 3000 different items?

All classification examples I see are using less then 10 classes.

(I saw lessons 1-4 so far)

2

Wow, 3000 classes are in fact a lot. But I think Imagenet also contains 1000 classes of object localization.

An error rate of 60% sounds bad by itself, but if you think about the fact that a random draw would only have a 1/3000 \approx 0.00033 chance of success, an accuracy of 40% is actually around 1200 times better. So I wouldn’t say the classifier is that bad :slight_smile: It’s simply a very hard classification problem given only very few examples to learn from.

So what I would do is to decrease the number of classes, e.g. at first to 1000, and see how the error rate changes and then try to add more training data.

3

So it should be possible do identify like 100 games and then go up. But I need get more training data.
I wasn’t sure whether its worth spending time on it, but know I it is :D.

Thanks a lot.

4

Well, no guarantees that it’ll work. I have no experience with training classifiers with such a high number of classes. I only wanted to offer a more optimistic perspective on your error rate and suggest to slowly and iteratively try to increase training set size :slight_smile: But in the end only you can decide a) how good your model needs to be for it to be considered a success and b) if it’s worth spending your time on it.

Good luck and keep us updated :wink:

5

you may check Kaggle Landmark Recognition and Retrieval Competitions [link] It deals with huge number of classed (https://www.kaggle.com/c/landmark-recognition-2019/discussion/95176) . related paper:


code not working properly but should be there in https://github.com/PaddlePaddle/
https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/Research/landmark

1 Like