research-ms-loss's Introduction

Multi-Similarity Loss for Deep Metric Learning (MS-Loss)

Code for the CVPR 2019 paper Multi-Similarity Loss with General Pair Weighting for Deep Metric Learning

Performance compared with SOTA methods on CUB-200-2011

Rank@K	1	2	4	8	16	32
Clustering⁶⁴	48.2	61.4	71.8	81.9	-	-
ProxyNCA⁶⁴	49.2	61.9	67.9	72.4	-	-
Smart Mining⁶⁴	49.8	62.3	74.1	83.3	-
Our MS-Loss⁶⁴	57.4	69.8	80.0	87.8	93.2	96.4
HTL⁵¹²	57.1	68.8	78.7	86.5	92.5	95.5
ABIER⁵¹²	57.5	68.7	78.3	86.2	91.9	95.5
Our MS-Loss⁵¹²	65.7	77.0	86.3	91.2	95.0	97.3

Prepare the data and the pretrained model

The following script will prepare the CUB dataset for training by downloading to the ./resource/datasets/ folder; which will then build the data list (train.txt test.txt):

./scripts/prepare_cub.sh

Download the imagenet pretrained model of bninception and put it in the folder: ~/.torch/models/.

Installation

pip install -r requirements.txt
python setup.py develop build

Train and Test on CUB200-2011 with MS-Loss

./scripts/run_cub.sh

Trained models will be saved in the ./output/ folder if using the default config.

Best recall@1 higher than 66 (65.7 in the paper).

Contact

For any questions, please feel free to reach

[email protected]

Citation

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={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
pages={5022--5030},
year={2019}
}

License

MS-Loss is CC-BY-NC 4.0 licensed, as found in the LICENSE file. It is released for academic research / non-commercial use only. If you wish to use for commercial purposes, please contact [email protected].

Recommend Projects