Mining on Manifolds: Metric Learning Without Labels

This is a Matlab/Python package for our paper:

A. Iscen, G. Tolias, Y. Avrithis, O. Chum. "Mining on Manifolds: Metric Learning Without Labels", CVPR 2018

It implements unsupervised selection of training pairs (in MATLAB) and training for fine-grained categorization (in Python/PyTorch).

Prerequisites

1. Package for diffusion proposed in our CVPR17 paper (automatically downloaded):

A. Iscen, G. Tolias, Y. Avrithis, T. Furon, O. Chum. "Efficient Diffusion on Region Manifolds: Recovering Small Objects with Compact CNN Representations", CVPR 2017

2. MatConvNet:

It is used to extract descriptors for training images using a pre-trained network. This will be the input to the mining process. The code is tested with MatConvNet version 1.0-beta25 in MATLAB R2016b.

3. Pytorch:

It is used to implement the training for fine-grained categorization on the CUB-200-2011 dataset. The code is tested with PyTorch version 1.0.1.post2 and Python 2.7.13 on Debian 8.1.

Execution

Demo on 1M web-crawled images

We provide the initial descriptors for 1 million images (particular object retrieval experiment in CVPR18), perform training pair selection with MoM, and qualitatively present results using image thumbnails.

Run the following script through MATLAB:

>> run mat/mom_1M

Fine-grained categorization

Training, in the form of metric learning, is performed for fine-grained bird categorization. The initial descriptor extraction and selection by MoM is performed in MATLAB. The training, based on an earlier version of this package, is performed in Python/Pytorch.

Download CUB_200_2011 with:

python py/download_dataset.py

Extract descriptors for training images by running in MATLAB:

>> run mat/extract_descriptors_CUB

Perform data selection for training by running in MATLAB:

>> run mat/mom_CUB

Perform the training with:

>> python py/train.py