Note that Metric Learning is also known as:

  • Similarity learning
  • Contrastive Learning
  • Embedding learning

Metric learning is the task of training a similarity function that measures how similar or related two (or 3) objects are.

Metric learning is very useful for classifications problems when the data is unbalanced or there are very few samples.

The input data becomes from 1 sample (classification) to 2 samples (metric learning) by using Siamese Networks

This augment your dataset a lot (Combinations of N objets in pairs formula). If you have 100 images, your net sees 4950 cases!

Clasification VS Metric learning

Classification Metric learning
Embeddings from different classes need to be easily separable. Embedding from the same class need to be close together, and embedding from different classes need to be far from each others.

Mining

Mining is the process of finding the best pairs or triplets to train on. T

Clasification Losses

  • Categorical Cross-Entropy Loss,
  • Binary Cross-Entropy Loss,
  • Softmax Loss,
  • Logistic Loss,
  • Focal Loss

Metric Learning Losses

Cross-entropy is valid choice of loss function, same as contrastive or triplet on L2 distance.

  • Pairwise Losses (Parejas de datos de entrada)
    • Concat embs -> Binary classification (fastai siamese)
      • Downside: The model is not “symmetrical”, (the input of [img A, img B] will give a different output from [img B, img A]
    • Absolute difference of embs -> Binary classification
    • Cosine similarity of embs -> Regression
  • Triplet Loss (Trios como datos de entrada)
    1. Anchor: represents a reference
    2. Positive: same class as the anchor
    3. Negative: a different class
  • Contrastive Loss
    • ArcFace
    • CosFace
    • SphereFace
  • Margin Loss
  • Hinge Loss

Kaggle competitions

Competition Reward End date Teams Category Classes Eval metric
Landmark Recognition 2021 0$ 2021-10-01 392 Research 81.313 GoogleGlobalAP
Landmark Retrieval 2021 0$ 2021-10-01 268 Research 81.313 mAP@100
Hotel-ID 2021 0$ 2021-05-26 92 Research 7.770 Mean AP at K
Herbarium 2021 0$ 2021-05-26 80 Research 64.500 macro F1 score
Shopee - Pruduct matching 30,000$ 2021-05-10 2464 Featured 11.014 Mean F1 score
Landmark Recognition 2020 25,000$ 2020-09-29 763 Research 81.313 GlobalAP (GAP)
Landmark Retrieval 2020 25,000$ 2020-08-17 544 Research 81.313 mAP@100
Herbarium 2020 0$ 2020-05-26 157 Research    
Herbarium 2019 0$ 2019-06-07 21 Community    
Landmark Recognition 2019 25,000$ 2019-06-03 281 Research    
Landmark Retrieval 2019 25,000$ 2019-06-03 144 Research    
Landmark Recognition 2018 2,500$ 2018-05-29 209 Research    
Landmark Retrieval 2018 2,500$ 2018-05-29 209 Research    

SimCLR: Metric Learning for unsupervised data

Reference