jenicek/mdir

Multi-Domain Image Retrieval

Multi-Domain Image Retrieval

A code repository for the following publication:

No Fear of the Dark: Image Retrieval under Varying Illumination Conditions
Tomas Jenicek and Ondřej Chum
In International Conference on Computer Vision (ICCV), 2019

Related:  project website,  paper pdf

This codebase builds on top of cirtorch. Its patched version is distributed together with the code and its dataset format is honored.

Getting Started

Clone repository, install dependencies:

git clone https://github.com/jenicek/mdir.git
cd mdir
pip3 install -r requirements.txt

Optionally, two environment variables may be specified:

• CIRTORCH_ROOT -- path for downloaded datasets, default is mdir top folder
• CUDA_VISIBLE_DEVICES -- gpu to be used for inference (training), default is index 0

Test inference:

cd mdir/examples/iccv19
./eval.py test

Datasets will be downloaded and stored in CIRTORCH_ROOT, trained models will be loaded from remote urls each time (see yaml scenario files). You should get following output:

    roxford.5k medium    39.06
    rparis.6k medium     58.94
    247tokyo.1k          72.0

The test argument is a shortcut for a yaml scenario which fully defines the evaluation. If multiple yaml scenarios are provided, they will be overlayed in the order they were provided. Shortcut test is equal to arguments eval.yml eval_test.yml.

Evaluation

In order to evaluate trained models from the ICCV19 paper, the following scenarios are provided in mdir/examples/iccv19

• eval_clahe.yml (shortcut clahe) -- "CLAHE N/D" method
• eval_composition.yml (shortcut composition) -- "U-Net jointly N/D" method

Configuration common for both scenarios is in eval.yml

Training

Currently, convenient scripts are provided only for evaluation. For training, only the code is provided. The repository is under development; training scripts will appear soon.

mdir's Issues

Questions about CLAHE N/D

Hi, I have two questions:

1. you mentioned that the best performing method is the VGG GeM combined with CLAHE N/D, so have you compared your retrieval performance with NetVlad? Which one is better?

2. Can you provide the test script then I can used to test the retrieval performance?

Really really hope to get your reply. Thanks in advance and have a nice day ! :)