2
$\begingroup$

I am working on a problem that requires the classification of more than 60k classes. I have around 1k to 1.5k images per class. I am using synthetic data for training and want to evaluate it on real data. I have enough computing power but want to keep it computationally efficient and highly accurate (the tradeoff can be further adjusted).

Currently, I am looking for papers in this direction. All papers mostly work with ImageNet 1k. I have a few things in my mind. I am considering starting with EfficientNet for supervised learning. I am also looking into Hierarchical classification and similarity matching by generating embeddings in multidimensional space.

The data does not have a hierarchy. But I am also looking into it if I could somehow use it in hierarchies.

I want suggestions on this. What methodology is best for it? or if there are any good papers.

$\endgroup$

2 Answers 2

1
$\begingroup$

Look at CLIP, it's somewhat relevant. The model can is capable of classifying a non-fixed number of classes

$\endgroup$
2
  • $\begingroup$ I thought so as well. If you use an embedding model you can always use kNN for classification $\endgroup$
    – Ggjj11
    Mar 24 at 18:36
  • $\begingroup$ I will use CLIP next currently working with pretrained models and metric learning. I am getting issue with embeddings because I have classes which are very similar and they don't have proper clusters. So, classification using kNN won't work. @Ggjj11 $\endgroup$
    – pks
    Apr 19 at 6:59
0
$\begingroup$

If you have a lot of classes that can be very similar (or not), you could maybe take a look at margin losses such as Arcface of Cosface.

Also start from a very well pretrained foundation model such as DINOv2 and fit a small classifier (even a simple linear layer could work decently) on top of its features.

$\endgroup$
1
  • $\begingroup$ Yes, I am working with pretrained models for classification and also looking into margin losses such as Triplet loss. $\endgroup$
    – pks
    Apr 19 at 6:55

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .