This work is being extended for TPAMI submission, with the main target to improve this work further.
- The updated codes are released. Feel free to have a try!
- In Unbiased Evaluation of Deep Metric Learning Algorithms--Istvan Feh ´ erv ´ ari etal 2019, it stated "On the SOP dataset, we never managed to make this algorithm converge." using Ranked List Loss.
- First of all, I thank their interest in our work, which is inspiring for me and my collaborators.
- However, their statement is not the fact. Please check our reproducible results.
- Paper Summary on Distance Metric, Representation Learning-Updated Blog
Recently, in paper A Metric Learning Reality Check, it is reported that the improvements over time have been marginal at best. Is it true? I present my personal viewpoints as follows:
-
First of all, acedemic research progress is naturally slow, continuous and tortuous. Beyond, it is full of flaws on its progress. For example,
- In person re-identification, several years ago, some researchers vertically split one image into several parts for alignment, which is against the design of CNNs and non-meaningful. Because deep CNNs are designed to be invariant against translation, so that hand-crafted alignment is unnecessary.
- The Adam optimiser is found to be very sensitive to the setting of delta recently.
- ...
-
There are some vital breakthroughs over the time although they seem to be trivial now.
-
Multibatch Method, after which people rarely use rigid input formats.
- Before this milestone, we heard a lot about siamese networks, triplet networks, etc.
- After Multibatch Method, we construct doublets, triplets, or high-order tuples directly in the embedding space.
-
The importance of sample mining/weighting becomes clearer for our community. Of course, there exist many variants of sample mining/weighting for different scenarios.
-
Our community become much more open-minded: all methods which learn discriminative representations can be categorised into deep metric learning, e.g., softmax + categorical cross entropy.
-
...
-
If you find our code and paper help your research, please kindly cite our work:
InProceedings{Wang_2019_CVPR,
author = {Wang, Xinshao and Hua, Yang and Kodirov, Elyor and Hu, Guosheng and Garnier, Romain and Robertson, Neil M.},
title = {Ranked List Loss for Deep Metric Learning},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2019}
}cd ./Ranked-List-Loss-for-Deep-Metric-Learning
tree
The core functions are implemented using C++ in the caffe framework. We use matlab interfaces matcaffe for data preparation.
-
Clone our repository: Simply copy and execute following commands in the command line
git clone [email protected]:XinshaoAmosWang/Ranked-List-Loss-for-D eep-Metric-Learning.git cd Ranked-List-Loss-for-Deep-Metric-Learning/
-
Install dependencies on Ubuntu 16.04
sudo apt-get install libprotobuf-dev libleveldb-dev libsnappy-dev libopencv-dev libhdf5-serial-dev protobuf-compiler sudo apt-get install --no-install-recommends libboost-all-dev sudo apt-get install libopenblas-dev sudo apt-get install python-dev sudo apt-get install libgflags-dev libgoogle-glog-dev liblmdb-dev
-
Install MATLAB 2017b
Download and Run the install binary file
./install
-
Compile Caffe and matlab interface
Note you may need to change some paths in Makefile.config according your system environment and MATLAB path
cd CaffeMex_RLL_GR_V03_Simp make -j8 && make matcaffe
-
Data preparation for SOP
Downlaod Stanford_Online_Products dataset from ftp://cs.stanford.edu/cs/cvgl/Stanford_Online_Products.zip
For simplicity, you can use the data mat file in pre_post_process directory, which is ready training and testing scripts. To solve the data path, you can do eithor a or b:
a. Changing the path within the mat files. b. A Simpler way: Create a soft link of your data e.g sudo ln -s /.../Stanford_Online_Products /home/xinshao/Papers_Projects/Data/Stanford_Online_Products -
Train & Test
Run the training and testing scripts in the training folder of a specific setting defined by its corresponding prototxt folder.
- Procedures are similar to SOP
- Please see training scripts in Folder InshopClothes_Simp_21042020
- ./SOP_Simp_25042020/pretrain_model/b150_v50_iter24k_0.748.caffemodel
- ./InshopClothes_Simp_21042020/train_60_T10_pn04_v63/checkpoints1/checkpoint_iter_26000.caffemodel
You only need to create training/testing mat files with the same structure as SOP_TrainImagePathBoxCell.mat and SOP_TestImagePathBoxCell.mat in directory SOP_GoogLeNet_Ori_V05/pre_pro_process.
e.g. SOP_TrainImagePathBoxCell.mat contains 
, TrainImagePathBoxCell storing all image paths and class_ids storing their corresponding semantic labels.
- Training scripts on Market1501 dataset are given in Folder Market1501_GooLeNetV2_V02_11042020;
- Scripts on data processing are in Folder Market-1501_baseline-v16.01.14
-
Given a query, the objective is to rank its postive set in front of its negative set by a distance margin.
-
We do not need to consider the exact order of examples within the positive and negative sets.
Please see our discussion in the paper.
Our work benefits from:
-
Hyun Oh Song, Yu Xiang, Stefanie Jegelka and Silvio Savarese. Deep Metric Learning via Lifted Structured Feature Embedding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. http://cvgl.stanford.edu/projects/lifted_struct/
-
CaffeMex_v2 library: https://github.com/sciencefans/CaffeMex_v2/tree/9bab8d2aaa2dbc448fd7123c98d225c680b066e4
-
Caffe library: https://caffe.berkeleyvision.org/
BSD 3-Clause "New" or "Revised" License
Xinshao Wang (You can call me Amos as well) xinshao dot wang at eng dot ox dot ac dot uk