AvitoTech

/

DINO-v2-small-for-animal-identification

like 1

Follow

AvitoTech 97

@drzraf
Thank you for your comment. We noticed the issue and fixed it. 🤝🔧

We've updated the description of how the model should be initialized in the README.
It should work better now, so, you may want to check this out – we expect better results!

Wow, it's way way better. I could reduce the similarity to 0.85 threshold and got, if not perfect (a couple of misses) at least very decent results !

Would you mind providing some more information model outputs. There structures, how to best used them, visualized(?) and consider from a high-level perspective?

Another questions, the models takes a list of PIL Image, but passing multiple arguments only returns one tensor. Am I doing something wrong? (I didn't tried a Dataset yet, but I believe it would have to work with just plain list to begin with).

Big thanks to your team!
Keep up the good work!

ref	match_1	match_2	match_3	match_4	match_5
	OK 0.982	OK 0.977	OK 0.963	OK 0.937	OK 0.911
	OK 0.951	OK 0.912	OK 0.908	OK 0.893
	OK 0.907
	OK 0.928	OK 0.905	OK 0.903	OK 0.874
	OK 0.975	OK 0.950	OK 0.899
	OK 0.889
	OK 0.930	OK 0.928	OK 0.909	OK 0.904
	OK 0.930	OK 0.925	OK 0.903	OK 0.885
	OK 0.936	OK 0.932	OK 0.906	OK 0.888
	OK 0.971	OK 0.966	OK 0.951	OK 0.883
	OK 0.948	OK 0.906	OK 0.893	OK 0.873
	OK 0.951	OK 0.947	OK 0.941	OK 0.921
	OK 0.932	OK 0.924	OK 0.915	OK 0.911	OK 0.908
	OK 0.934	OK 0.876	OK 0.868
	OK 0.921	OK 0.911	OK 0.909	OK 0.883
	OK 0.924	OK 0.912	OK 0.906	OK 0.894
	OK 0.881
	OK 0.931	OK 0.894
	OK 0.907
	OK 0.937	OK 0.932	OK 0.932	OK 0.930	OK 0.917
	OK 0.931	OK 0.916
	OK 0.932	OK 0.925	OK 0.864	OK 0.862
	OK 0.909	OK 0.902	OK 0.885	OK 0.874
	OK 1.000	OK 0.909
	OK 0.883	OK 0.871
	OK 0.951	OK 0.924	OK 0.908	OK 0.873
	OK 0.932	OK 0.912	OK 0.900
	OK 0.904
	OK 1.000	OK 0.909
	OK 0.877	OK 0.853
	OK 0.903	OK 0.853
	OK 0.934	OK 0.916	OK 0.873
	OK 0.928	OK 0.877
	OK 0.967	OK 0.965	OK 0.906	OK 0.870
	OK 0.948	OK 0.908	OK 0.908	OK 0.894
	OK 0.977	OK 0.969	OK 0.957	OK 0.930	OK 0.924
	OK 0.916	OK 0.894
	OK 0.934	OK 0.916	OK 0.868
	OK 0.952	OK 0.950
	OK 0.899	OK 0.891
	OK 0.992	OK 0.966	OK 0.941	OK 0.909
	OK 0.967	OK 0.933	OK 0.888
	OK 0.904
	OK 0.928	OK 0.889	OK 0.881	OK 0.853
	OK 0.863
	OK 0.932	OK 0.900	OK 0.883	OK 0.870
	OK 0.937	OK 0.853
	OK 0.969	OK 0.963	OK 0.935	OK 0.917	OK 0.908
	OK 0.992	OK 0.971	OK 0.947	OK 0.911
	OK 0.912	OK 0.864	OK 0.862
	OK 0.909	OK 0.909
	OK 0.965	OK 0.936	OK 0.933	OK 0.900	OK 0.871
	OK 0.975	OK 0.952	OK 0.891	OK 0.870
	OK 0.937	OK 0.903
	OK 0.870
	OK 0.925	OK 0.905	OK 0.904	OK 0.902
	OK 0.936	OK 0.864
	OK 0.934	OK 0.876	OK 0.873
	OK 0.936	OK 0.925	OK 0.900	OK 0.864
	OK 0.863
	OK 0.982	OK 0.957	OK 0.935	OK 0.932	OK 0.915

@drzraf

Thanks a lot for the detailed feedback, it is great to hear that updated version gives you very reasonable results.

Regarding the model outputs: we discuss their structure, interpretation, and recommended ways to use them (including some visualization ideas) in a paper that is currently under review. As soon as the paper is published, we will add a link to it in the repository so that all the details are documented in one place.

About your second question: the behavior you see is expected. The model takes a list of PIL.Image objects, internally stacks them into a batch, and returns a single tensor of shape [batch_size, embedding_dim] where batch_size is the number of images in your list, which is the usual convention for PyTorch vision models. You may not have to use a Dataset for simple experiments; passing a plain list is perfectly fine. Here is a minimal example you can use to process a set of images:

import glob
from PIL import Image
import torch
import torch.nn.functional as F

paths = glob.glob("*.jpeg")
images = [Image.open(p).convert("RGB") for p in paths]

with torch.no_grad():
    embeddings = model(images)          # shape: [len(images), embedding_dim]
    embeddings = F.normalize(embeddings, dim=1)

print(embeddings.shape)

Thanks again for the kind words and for taking the time to test the model so thoroughly. This kind of feedback is very helpful for us.