ISC21-Descriptor-Track-1st

The 1st Place Solution of the Facebook AI Image Similarity Challenge (ISC21) : Descriptor Track.

You can check our solution tech report from: Contrastive Learning with Large Memory Bank and Negative Embedding Subtraction for Accurate Copy Detection

Main features:

• The weights of the competition winning models are publicly available and easy to use.
• Without any fine-tuning or something, our models work well with image/video copy detection, image retrieval, and so on.
  • In video copy detection task, it is reported that our model has the best result among recent frame feature extractor, despite with the smallest feature dimensionality (ref: https://github.com/alipay/VCSL).

Installation

pip install git+https://github.com/lyakaap/ISC21-Descriptor-Track-1st

Usage

import requests
import torch
from PIL import Image

from isc_feature_extractor import create_model

recommended_weight_name = 'isc_ft_v107'
model, preprocessor = create_model(weight_name=recommended_weight_name, device='cpu')

url = "http://images.cocodataset.org/val2017/000000039769.jpg"
image = Image.open(requests.get(url, stream=True).raw)
x = preprocessor(image).unsqueeze(0)

y = model(x)
print(y.shape)  # => torch.Size([1, 256])