Skip to content

Source code for the paper "Divide and Conquer the Embedding Space for Metric Learning", CVPR 2019

License

Notifications You must be signed in to change notification settings

CompVis/metric-learning-divide-and-conquer

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

8 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Divide and Conquer the Embedding Space for Metric Learning

About

This repository contains the code for reproducing the results for Divide and Conquer the Embedding Space for Metric Learning (CVPR 2019) with the datasets In-Shop Clothes, Stanford Online Products and PKU VehicleID.

Paper: pdf
Supplementary: pdf

We also applied our method to the Humpback Whale Identification Challenge at Kaggle and finished at 10th place out of 2131.
Slides: link

method pipeline

Requirements

Usage

The following command will train the model with Margin loss on the In-Shop Clothes dataset for 200 epochs and a batch size of 80 while splitting the embedding layer with 8 clusters and finetuning the model from epoch 190 on. You can use this command to reproduce the results of the paper for the three datasets by changing simply --dataset=inshop to --dataset=sop (Stanford Online Products) or --dataset=vid (Vehicle-ID).

CUDA_VISIBLE_DEVICES=0 python experiment.py --dataset=inshop \
--dir=test --exp=0 --random-seed=0 --nb-clusters=8 --nb-epochs=200 \
--sz-batch=80 --backend=faiss-gpu  --embedding-lr=1e-5 --embedding-wd=1e-4 \
--backbone-lr=1e-5 --backbone-wd=1e-4 --finetune-epoch=190

The model can be trained without the proposed method by setting the number of clusters to 1 with --nb-clusters=1.
For faster clustering we run Faiss on GPU. If you installed Faiss without GPU support use flag --backend=faiss.

Expected Results

The model checkpoints and log files are saved in the selected log-directory. You can print a summary of the results with python browse_results <log path>.

You will get slightly higher results than what we have reported in the paper. For SOP, In-Shop and Vehicle-ID the R@1 results should be somewhat around 76.40, 87.36 and 91.54.

Related Repos

  • Collection of baselines for metric learning from @Confusezius [PyTorch]

License

You may find out more about the license here

Reference

If you use this code, please cite the following paper:

Artsiom Sanakoyeu, Vadim Tschernezki, Uta Büchler, Björn Ommer. "Divide and Conquer the Embedding Space for Metric Learning", CVPR 2019.

@InProceedings{dcesml,
  title={Divide and Conquer the Embedding Space for Metric Learning},
  author={Sanakoyeu, Artsiom and Tschernezki, Vadim and B\"uchler, Uta and Ommer, Bj\"orn},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  year={2019},
}

About

Source code for the paper "Divide and Conquer the Embedding Space for Metric Learning", CVPR 2019

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages