Deep Metric Learning Beyond Binary Supervision

outline

Official pytorch Implementation of Deep Metric Learning Beyond Binary Supervision, CVPR 2019

Citing this work

If you find this work useful in your research, please consider citing:

@inproceedings{kim2019deep,
  title={Deep Metric Learning Beyond Binary Supervision},
  author={Kim, Sungyeon and Seo, Minkyo and Laptev, Ivan and Cho, Minsu and Kwak, Suha},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={2288--2297},
  year={2019}
}

Dependency

Python >=3.6
Pytorch >=0.4.1
tqdm (pip install tqdm)
scipy
tensorboardX

Prerequisites

Download pretrained human pose dataset with labels from here
Extract the zip file into ./data/

Human Pose Retrieval Quick Start


python main.py --help

# Train a embedding network of resnet34 (d=128)
# using logratio loss with **dense triplet sampling**.

python main.py --loss logratio \
               --model resnet34 \ 
               --result-name dense_Logratio \
               --optimizer sgd \
               --lr 0.01 \ 
               --lr-decay 1e-4 \ 
               --batch-size 150 \
               --num-NN 5 \
               --embedding-size 128 \
               --sampling dense \

# Train a embedding network of resnet34 (d=128)
# using triplet loss (margin=0.03) with **dense triplet sampling**.

python main.py --loss triplet \
               --is-norm True \
               --model resnet34 \ 
               --result-name dense_Triplet \
               --optimizer sgd \
               --lr 0.01 \ 
               --lr-decay 1e-4 \ 
               --batch-size 150 \
               --num-NN 5 \
               --embedding-size 128 \
               --sampling dense \               

# Train a embedding network of resnet34 (d=128)
# using triplet loss (margin=0.2) with **binary triplet sampling**.

python main.py --loss triplet \
               --is-norm True \
               --model resnet34 \ 
               --result-name naive_Triplet \
               --optimizer sgd \
               --lr 0.01 \ 
               --lr-decay 1e-4 \ 
               --batch-size 150 \
               --embedding-size 128 \
               --sampling naive \