docs

Running the baseline code of the ISC2021 challenge

This is a step-by-step walkthrough on how to get a valid submission for the ISC2021 challenge.

This tutorial will guide you to get baseline results for the ISC2021 competion, how to evaluate them and get a submittable file.

Downloading the data

Get the images for query, reference and training sets as described in the Driven Data page

https://www.drivendata.org/competitions/80/competition-image-similarity-2-dev/data/

Please be patient, this is a total of 350 GB of data. Note that the training images are not required for the first steps of the process.

Update (2022-02-08): After the competition, the data is available at: https://sites.google.com/view/isc2021/dataset

In the following, we assume that the images are available in the images/queries, images/references and images/train subdirectories.

While the data is downloading, you can install the required packages and compile some code.

Cloning & installing dependencies

First, clone this repo:

git clone https://github.com/facebookresearch/isc2021.git

Follow the steps below to install all the required dependencies in order to run the ISC evaluation code. Note: The code in this repo requires 3.5 <= Python <= 3.8.

conda create -n isc2021 python=3.8 -y && conda activate isc2021
pip install -e isc2021/
conda install -c pytorch faiss-gpu

Steps

The tutorial breaks down in steps from easiest and fastest to more complicated.

Step 1: GIST descriptors on a small subset

Step 2: GIST descriptors with PCA

Step 3: GIST descriptors on the full dataset

Step 4: Multigrain descriptors on the small subset

Step 5: Multigrain descriptors on the full dataset

Step 6: HOW local features on the small subset

Step 7: HOW local features on the full dataset