POVa 2024 Project - Content-Based Image Search

Installation

1. Install the required packages:

pip install -r requirements.txt

2. Create required directories:

mkdir -p models
mkdir -p images
mkdir -p indices

Usage

Setting up the Dataset

1. Prepare the COCO dataset with embeddings using the chosen model:

python create_indices.py --use-coco -m CLIP

The -m flag specifies which model to use. Available options are:

• CLIP (default ViT-B/32)
• CLIP_L (ViT-L/14)
• DINOv2
• ResNet
• ViT
• dh_vit - DeepHash ViT
• dh_resnet - DeepHash ResNet

Running Benchmarks

1. Performance benchmarks:

python benchmark.py --db_size 1000000 --search_size 20000 indices/clip.npz

2. Memory usage analysis:

python benchmark_memory.py --db_size 1000000 indices/clip.npz

3. Indexing method benchmarks:

# FAISS benchmarks
python benchmark_faiss.py indices/clip.npz

# HNSW benchmarks
python benchmark_hnsw.py indices/clip.npz

Running the Search Interface

Start the web interface:

python app.py