clovaai / symmetrical-synthesis

Official Tensorflow implementation of "Symmetrical Synthesis for Deep Metric Learning" (AAAI 2020)

Symmetrical Synthesis for Deep Metric Learning

Official Tensorflow implementation of Symmetrical Synthesis for Deep Metric Learning (AAAI 2020)

Geonmo Gu*, Byungsoo Ko* (* Authors contributed equally.)

@NAVER/LINE Vision

• Paper | Supplementary Video | Poster | Spotlight PPT

Overview

Symmetrical Synthesis

Symmetrical Synthesis (Symm) is a novel method of synthetic hard sample generation for deep metric learning.

How it Works

After some iterations, symmetrical synthesis generates synthetic points around the class clusters, which are used as hard samples to push the other class with stronger power.

Experimental Results

Getting Started

Requirements

$ pip3 install -r requirements.txt

Prepare Data

1. Download pretrained GoogleNet model. ref

$ wget https://github.com/Wei2624/Feature_Embed_GoogLeNet/raw/master/tf_ckpt_from_caffe.mat

2. Download CAR DB and cook.

$ wget http://ai.stanford.edu/~jkrause/car196/car_ims.tgz
$ tar -xzf car_ims.tgz
$ mv car_ims
$ wget http://ai.stanford.edu/~jkrause/car196/cars_annos.mat

# on ../symm_public folder
$ cd dataset
$ python3 cooking_CARS.py --car_folder=/your/car_ims/folder \
--save_path=/your/converted/carDB/will/be/saved/here

Train a Model

• Available losses: N-pair, Symm + N-pair, Angular, Symm + Angular

Symm + N-pair

$ python3 train.py --backbone=googlenet \
--pretrained_model_path=/your/folder/tf_ckpt_from_caffe.mat \
--image_path=/your/converted/carDB/will/be/saved/here \
--run_gpu=0 \
--save_path=/your/trained/model/will/be/saved/here \
--losses=symm_npair --dim_features=512 \
--input_size=227 --learning_rate=0.0001 \
--decay_steps=5000 --decay_stop_steps=15000 \
--decay_stop_value=0.00001 --decay_ratio=0.5 \
--save_model_steps=100

Test a Model

$ python3 test.py --run_gpu=1 --model_path=/your/trained/model/will/be/saved/here \
--image_path=/your/converted/carDB/will/be/saved/here \
--batch_size=512 --backbone=googlenet \
--pretrained_model_path=/your/folder/tf_ckpt_from_caffe.mat \
--log_path=eval_log_car \
--input_size=227 --start_idx=0 --dim_features=512

• Best recall@1: 0.77 (0.765 in paper)

Check Test Results

$ tensorboard --logdir=eval_log_car --port=10000

Acknowledgements

• Googlenet backbone
  • https://github.com/wzzheng/HDML/blob/master/lib/GoogleNet_Model.py
  • https://github.com/Wei2624/Feature_Embed_GoogLeNet

Citation

If you find Symmetrical Synthesis useful in your research, please consider to cite the following paper.

@inproceedings{gu2020symmetrical,
    title={Symmetrical Synthesis for Deep Metric Learning},
    author={Geonmo Gu and Byungsoo Ko},
    booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
    year={2020}
}

License

Copyright (c) 2020-present NAVER Corp.

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in
all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
THE SOFTWARE.

symmetrical-synthesis's Issues

embedding network architecture

After average pooling the CNN features extracted from the googlenet,

It looks like 'slim.conv2d' was used instead of a fully connected layer.

Then the code in this repository has a different architecture from the original embedding network.

Is this correct?