UCU homework - SOP_Metric_Learning
You are provided with Stanford Online Products dataset [DL]. It contains train (Ebay_train) and test (Ebay_test) sets.
- Train set should be divided into train and val.
You need to choose a backbone [one] from torchvision.models and perform:
- Build an index based on the training set and perform the retrieval of the val set using any library you want [faiss, annoy, etc].
- Estimate Accuracy and mAP5 for both class and super_class for the val set.
- Perform the retrieval for the test set.
- For each class choose a 3-5 pictures and generate visualizations similar to [slide]
Repeat previous actions for:
- Plain pre-trained ImageNet backbone
- Fine-tuned with vanilla Cross-Entropy and classification approach
- Fine-tuned using ArcFace Loss
- Fine-tuned using the Siamese approach and Contrastive Loss
- [optional] Fine-tuned using Triplet Loss