Metric Learning

Metric learning or Similarity learning is the process of training a feature vector that can be used to acess its similarity between object classes. It can be used in ranking, recomendation systems, visual identity tracking, face verification and others.

Right now I found two diferent groups for this aproaches, contrastive and softmax.

Contrastive

1. Contrastive loss
2. Triplet loss

Softmax

1. ArcFace
2. SphereFace
3. CosFace

ReId models and tracking

• OSNet ReID models
• Fast-ReId

Links

• PyTorch Metric Learning: An opinionated review
• OpenSphere