DINOv2

DINOv2

• Automatic pipeline to build a dedicated, diverse, and curated image dataset instead of uncurated data
• Focus on text-guided pretraining
• Textual supervision to guide the training of the features
• PCA between the patches of the images from the same column

• Features are learned from images alone
• Self-supervised learning has the potential to learn all-purposed visual features if pretrained on a large quantity of curated data
• Automatic pipeline to filter and rebalance datasets from an extensive collection of uncurated images
• Data similarities are used instead of external metadata and do not require manual annotation

Other Approaches

• Extracting a signal from the image to be predicted from the rest of the image
• Discriminative signals between images or groups of images to learn features. 

• Copy detection pipeline of Pizzi et al. (2022) to the uncurated data and remove near-duplicate images
• Compute an image embedding using a self-supervised ViT-H/16 network pretrained on ImageNet-22k, and use cosine-similarity as a distance measure between images.
• k-means clustering of the uncurated data.
• Query dataset for retrieval, if it is large enough we retrieve N (typically 4) nearest neighbors for each query image.

Summary

• DINOv2, a new series of image encoders pretrained on large curated data with no supervision
• Visual features are compatible with classifiers as simple as linear layers - meaning the underlying information is readily available

Keep Exploring!!!