Learn deep metric for image retrieval or other information retrieval. 

我写了一个知乎文章，通俗快速解读了XBM想法动机：

跨越时空的难样本挖掘

欢迎大家阅读指点！

Recommend one recently released excellent papers in DML not written by me:

from Cornell Tech and Facebook AI

Abstract: Deep metric learning papers from the past four years have consistently claimed great advances in accuracy, often more than doubling the performance of decade-old methods. In this paper, we take a closer look at the field to see if this is actually true. We find flaws in the experimental setup of these papers, and propose a new way to evaluate metric learning algorithms. Finally, we present experimental results that show that the improvements over time have been marginal at best.

• Great Improvement: XBM can improve the R@1 by 12~25% on three large-scale datasets

• Easy to implement: with only several lines of codes

• Memory efficient: with less than 1GB for large-scale datasets

• Code has already been released: xbm

pytorch-metric-learning(a great work by Kevin Musgrave)

• code released link

• New Version of paper , To make my idea to be understand easily, I have rewritten the major part of my paper recently to make it clear. (at 2020-03-24)

• Contrasstive Loss [1]

• Semi-Hard Mining Strategy [2]

• Lifted Structure Loss* [3] (Modified version because of its original weak performance)

• Binomial BinDeviance Loss [4]

• NCA Loss [6]

• Multi-Similarity Loss [7]

• Car-196

  first 98 classes as train set and last 98 classes as test set

• CUB-200-2011

  first 100 classes as train set and last 100 classes as test set

• Stanford-Online-Products

  for the experiments, we split 59,551 images of 11,318 classes for training and 60,502 images of 11,316 classes for testing

• In-Shop-clothes-Retrieval

  For the In-Shop Clothes Retrieval dataset, 3,997 classes with 25,882 images for training. And the test set are partitioned to query set with 3,985 classes(14,218 images) and gallery set with 3,985 classes (12,612 images).

• Processed CUB and Cars196

  Extract code: inmj

  To easily reimplement the performance, I provide the processed datasets: CUB and Cars-196.

• Python >= 3.5
• PyTorch = 1.0

*** weight :***

sh run_train_00.sh

[Tensorflow] (by geonm)

[1] [R. Hadsell, S. Chopra, and Y. LeCun. Dimensionality reduction by learning an invariant mapping]

[2] [F. Schroff, D. Kalenichenko, and J. Philbin. Facenet: A unified embedding for face recognition and clustering. In CVPR, 2015.]

[3][H. Oh Song, Y. Xiang, S. Jegelka, and S. Savarese. Deep metric learning via lifted structured feature embedding. In CVPR, 2016.]

[4][D. Yi, Z. Lei, and S. Z. Li. Deep metric learning for practical person re-identification.]

[5][C. Wu, R. Manmatha, A. J. Smola, and P. Kr¨ahenb¨uhl. Sampling matters in deep embedding learning. ICCV, 2017.]

[6][R. Salakhutdinov and G. Hinton. Learning a nonlinear embedding by preserving class neighbourhood structure. In AISTATS, 2007.]

If you use this method or this code in your research, please cite as:

@inproceedings{wang2019multi,
title={Multi-Similarity Loss with General Pair Weighting for Deep Metric Learning},
author={Wang, Xun and Han, Xintong and Huang, Weilin and Dong, Dengke and Scott, Matthew R},
booktitle={CVPR},
year={2019}
}

@inproceedings{wang2020xbm,
title={Cross-Batch Memory for Embedding Learning},
author={Wang, Xun and Zhang, haozhi and Huang, Weilin and Scott, Matthew R},
booktitle={CVPR},
year={2020}
}