Self-Supervised FGIR Pre-Training with Adaptive Sampling

This repository implements the method described in the paper "Self-Supervised Pre-Training with Adaptive Sampling for Fine-Grained Image Retrieval". The project leverages self-supervised pre-training combined with an adaptive sampling strategy (ASS) to build a fine-grained image retrieval (FGIR) system.

Main Features

• Model Architecture: Integrates contrastive and generative learning with an Adaptive Sample Selector (ASS) that dynamically selects training samples based on difficulty.
• Experimental Setup: Supports training and evaluation on datasets such as CUB-200, Cars-196, SOP, and In-Shop.
• Code Structure: Detailed in the repository structure below.

Installation

We recommend creating a virtual environment using conda or virtualenv:

pip install -r requirements.txt

Training

Adjust the parameters in experiments/config.yaml as needed, then start training with:

python scripts/train.py --config experiments/config.yaml

Evaluation

After training, evaluate your model using:

python scripts/evaluate.py --config experiments/config.yaml --checkpoint path/to/model_checkpoint.pth

Repository Structure

• src/:Contains the model definitions, modules, dataset loader, and utility tools.
• scripts/: Contains training and evaluation scripts.
• docs/: Documentation for model architecture and usage.
• tests/: Unit tests for various modules.

Citation

If you find this project helpful for your work, please cite our paper:

Xiaoqing Li, Ya Wang, "Self-Supervised Pre-Training with Adaptive Sampling for Fine-Grained Image Retrieval", 2025.