https://github.com/lyakaap/ISC21-Descriptor-Track-1st

The 1st Place Solution of the Facebook AI Image Similarity Challenge (ISC21) : Descriptor Track.
https://github.com/lyakaap/ISC21-Descriptor-Track-1st

Last synced: about 1 month ago
JSON representation

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

• Host: GitHub
• URL: https://github.com/lyakaap/ISC21-Descriptor-Track-1st
• Owner: lyakaap
• License: mit
• Created: 2021-06-29T06:57:34.000Z (almost 3 years ago)
• Default Branch: master
• Last Pushed: 2023-11-08T03:40:28.000Z (8 months ago)
• Last Synced: 2024-02-10T15:42:59.014Z (4 months ago)
• Language: Python
• Homepage:
• Size: 24.2 MB
• Stars: 114
• Watchers: 4
• Forks: 17
• Open Issues: 3
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Lists

• awesome-stars - ISC21-Descriptor-Track-1st - The 1st Place Solution of the Facebook AI Image Similarity Challenge (ISC21) : Descriptor Track. (Python)
• awesome - lyakaap/ISC21-Descriptor-Track-1st - The 1st Place Solution of the Facebook AI Image Similarity Challenge (ISC21) : Descriptor Track. (Python)

README

        
# 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](https://arxiv.org/abs/2112.04323)

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

```python

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])

```