lightly VS dino

Another year of has passed, and we’ve seen exciting progress in research around self-supervised learning in computer vision. We’re very excited that some of the recent models such as Masked Autoencoders (MAE) or Masked Siamese Networks (MSN) have been added to our OSS framework.

If you’re interested in self-supervised learning and want to try it out yourself you can check out our open-source repository for self-supervised learning.

I assume that you have an image of the invoice. Then using a framework like https://github.com/lightly-ai/lightly with many implemented algorithms is the way to go. And after that step, with model-producing embeddings, you need to compare the embedding of a query with your known database and check if the distance is below some threshold. Of course, pipeline with checking the closest neighbor can be more complicated but I would start with sth really simple.

https://github.com/lightly-ai/lightly implements a lot of self supervised models, and had been available for a while.

modAL indeed has a similar goal of choosing the best subset of data to be labeled. However it has some notable differences:

modAL is built on scikit-learn which is also evident from the suggested workflow. Lightly on the other hand was specifically built for deep learning applications supporting active learning for classification but also object detection and semantic segmentation.

modAL provides uncertainty-based active learning. However, it has been shown that uncertainty-based AL fails at batch-wise AL for vision datasets and CNNs, see https://arxiv.org/abs/1708.00489. Furthermore it only works with an initially trained model and thus labeled dataset. Lightly offers self-supervised learning to learn high dimensional embeddings through its open-source package https://github.com/lightly-ai/lightly. They can be used through our API to choose a diverse subset. Optionally, this sampling can be combined with uncertainty-based AL.

You can easily train, embed, and upload a dataset using the lightly Python package.  First, we need to install the package. We recommend using pip for this. Make sure you're in a Python3.6+ environment. If you're on Windows you should create a conda environment.

I also implemented one based on the self attention maps from the DINO trained ViT’s. This worked pretty well when the attention maps were combined with some traditional computer vision to get bounding boxes. It seemed an ok compromise between domain specialization and location specificity. I did not try any saliency or gradient based methods as i was not sure on generalization and speed respectively. I know LAVIS has an implementation of grad cam and it seems to work well in the plug'n'play vqa.

You will probably need an unwieldy amount of data and compute to reproduce it, so your best option would be to use the pretrained models available on github.

[Full Explanation Post] [Arxiv] [Project Page]

Except he is actually talking about the new DINO model created by facebook that was released on friday. Which is a new approach to image transformers for unsupervised segmentation. Here's its github.

Code: https://github.com/facebookresearch/dino