CSC Hackathon 2023 - LUN track solution

Team GARCH (Andrii Yerko, Nikita Fordui, Andrii Shevtsov, Sofiia Shaposhnikova)

Second place solution for LUN track of CSC Hackathon 2023.

Task

The task for the hackathon was near duplicate image comparison. The model should take two same slightly changed images (different gamma, watermark, etc.) and output 1 if the images are the same, and 0 - if different.

Solution

Our team's solution solved this task and placed second on the private leaderboard of the competition. Our solution is based on different image hashing and keypoint detection algorithms and combining results into gradient boosting model. Presentation for the solution can be found here.

Repository structure

data folder contains all data splits and other .csv files used in development e.g. test submission format.

code folder contains all the code produced during the hackathon. All code files split into 5 groups based on the part of the solution they are connected to. Each group has its name prefix:

• 0 - data analysis and preprocessing
• 1 - hash and key points experiments
• 2 - neural networks experiments
• 3 - boosting experience
• 4 - evaluation and MVP

Same folder also contains some util code for smoother workflow.

Run

To reproduce the results of the projects, you can use virtual environment with requirements.txt file.

First start with cloning this GitHub repository

To create virtual environment:

python -m venv <venv_name>

To install all the necessary dependencies, first, activate the environment:

source <venv_name>/bin/activate

After that, to install the dependencies run:

pip install -r requirements.txt

All the described steps are described for Linux operating system.