srebuffi / lsd-clusters Goto Github PK

I haven’t fully grasped the paper, but the idea of having pair-wise new labels within the minibatch is, in fact, the limiting factor of the algo.
Most of the images can only be reduced to 512x512 without losing much information, and passing Image of size 512x512 with mini-batch over 32 will result in a computational overflow.
correct me If I am wrong?

The color is another feature that often results in bias, using rot-net we may reduce some geometrical bias but still, there is no mention of the color.

So is the paper just a proof of concept?

Hi, How can I train the model on my own dataset?

Hi, may I ask if you plan to add the rest of the datasets (in terms of RotNet pre-trained models and/or code)? Thank you!

Thanks for your great work.
I wonder how big ram you used in the clustering step?

There will be ~1TB ram usage in the KNN calculation for example. The PairNum consumes a lot of memory if I'm not wrong.