Image Retrieval System — Competition 2025

Description

This repository contains the code developed for the Competition 2025 of the Introduction to Machine Learning course. The goal of the competition is to implement an image retrieval system that, given a set of test images, retrieves the k most similar images from the training set.

We adopted an approach based on fine-tuning pretrained models and a shared retrieval module.

---

Project Structure

competition/
│
├── config/                  # YAML configuration files for each model
│   ├── clip_vitL14.yaml
│   ├── eva01_g14.yaml
│   └── resnet50.yaml
│
├── data/                    # Dataset folders
│   ├── train/            # Training images
│   └── test/             # Test images
│       ├── query/        # Query images for retrieval
│       └── gallery/      # Gallery images for retrieval
│
├── results/                 # JSON retrieval results for each model
│   ├── results_clip_vitL14_finetuned.json
│   ├── results_clip_vitL14.json
│   ├── results_eva01_g14.json
│   ├── results_resnet50_finetuned.json
│   └── results_resnet50.json
│
├── saved_models/            # Fine-tuned models (not included in public repo)
│
├── finetune_clip.py         # Fine-tuning script for CLIP ViT-L/14
├── finetune_eva01.py        # Fine-tuning script for EVA-01-G/14
├── finetune_resnet.py       # Fine-tuning script for ResNet50
│
├── retrieval.py             # Image retrieval pipeline (shared among models)
├── submit.py                # Script to generate submission JSON
│
├── requirements.txt         # Required libraries
└── README.md             # This file

---

Configuration

All fine-tuning and retrieval scripts rely on .yaml configuration files specifying:

• Base model
• Training dataset
• Training parameters (learning rate, batch size, etc.)
• Path to save the trained models

---

Models Used

• ResNet50 (torchvision)
• CLIP ViT-L/14 (HuggingFace)
• EVA-01-G/14 (OpenCLIP)

Resnet50 and CLIP Vit-L/14 models are fine-tuned separately using their respective YAML configuration.

---

Retrieval Pipeline

The retrieval pipeline is shared across all models:

1. Feature extraction from training and test images
2. Automatic feature saving for fast reuse
3. Distance computation between features (Cosine)
4. Selection of the k most similar images
5. Saving results in a JSON file, in the format required by the competition:

[
  {
    "filename": "test_image.png",
    "samples": ["train_image1.png", "train_image2.png", ...]
  }
]

---

Usage

Fine-tuning:

python finetune_resnet.py --config config/resnet50.yaml
python finetune_clip.py --config config/clip_vitL14.yaml
python finetune_eva01.py --config config/eva01_g14.yaml

Retrieval:

python retrieval.py --config config/resnet50.yaml

Submission:

python submit.py

---

Metrics

• Top-k Accuracy (official metric of the competition)

• Top-1 Accuracy (600 points): the correct identity is the first retrieved image.
• Top-5 Accuracy (300 points): the correct identity appears in the top-5 retrieved images.
• Top-10 Accuracy (100 points): the correct identity appears in the top-10 retrieved images

---

Requirements

• torch
• torchvision
• tqdm
• PyYAML
• scikit-learn
• numpy
• Pillow
• open_clip_torch
• optuna

Install everything with:

pip install -r requirements.txt