📌 Feature Extractor

This repository contains Python code to map images to representation vectors, which is useful for image retrieval.

Requirements

• Install the latest version of Python 3.X.
• Install the required packages:

pip install -r requirements.txt

Usage

Given a Github repository which includes hubconf.py, run:

!python extract_fts.py \
 --data_dir images --batch_size 256 \
 --resize_size 256 --keep_ratio --crop_size 224 \
 --model_repo "facebookresearch/dinov2" --model_name dinov2_vits14 \
 --output_dir features \
 --torch_features fts.pth \
 --numpy_features fts.npy \
 --img_list img_list.json \
 --verbose

The following files will be saved in the --output_dir folder:

• fts.pth: the features (as torch.float32) in a PyTorch file,
• fts.npy: the features (as np.float16) in a NumPy file,
• img_list.json: the list of image paths corresponding to the features.

Example

For instance, to extract features for images in the balloon dataset:

%cd /content
!git clone https://github.com/woctezuma/feature-extractor.git
%cd feature-extractor
%pip install --quiet -r requirements.txt

!wget https://github.com/matterport/Mask_RCNN/releases/download/v2.1/balloon_dataset.zip
!unzip -q balloon_dataset.zip

!python extract_fts.py --data_dir balloon

References

• A feature matcher based on the faiss library.
• match-steam-banners: retrieve games with similar store banners.
• steam-DINOv2: retrieve games with similar store banners, using Meta AI's DINOv2.
• Meta AI's Active Image Indexing:
  • Official blog post
  • Official Github repository
  • Fernandez, Pierre, et al. Active image indexing. ICLR 2023.