# Silver medal solution: ViT-L/14 with test time augmentations

Thank you to the host, organizers and all participants!
You can check out my solution here:
[https://www.kaggle.com/code/slavabarkov/google-universal-image-embedding-submission](https://www.kaggle.com/code/slavabarkov/google-universal-image-embedding-submission)
Brief summary of my solution:
ViT-L/14 encoder from the OpenCLIP implementation pre-trained on LAION-400M dataset, achieves 0.519 public/0.537 private if the input image is resized while maintaining the aspect ratio
Projection layer trained with ArcFace on the Products-10k dataset based on
[@motono0223](https://www.kaggle.com/motono0223)
work, achieves 0.598 public/0.608 private
Multiple test time augmentations, achieves 0.628 public/0.631 private:
Predefined crops - center crop and two crops with preselected offset
Pixel dropout - drops pixels from image according to predefined mask