Relational Self-supervised Distillation with Compact Descriptors for Image Copy Detection.

The official repository for Relational Self-supervised Distillation with Compact Descriptors for Image Copy Detection.

Pipeline

Implementation

• This code is implemented using Pytorch Lightning

Datasets used

• DISC (Facebook Image Similarity Challenge 2021)
• NDEC (A Benchmark and Asymmetrical-Similarity Learning for Practical Image Copy Detection 2023)
• CD10K(copydays + 10k distractors)

Performances

Method	Network	# of params[M]	Feature Size	µAPsn	download
DINO	ViT-B/16	85.8	1536	53.8
SSCD	ResNet-50	24.6	512	72.5
SSCD	EFF-B0	4.7	512	61.1
RDCD(ours)	EFF-B0	4.7	64	53.5	link
RDCD(ours)	EFF-B0	4.7	128	61.1	link
RDCD(ours)	Mobile-V3	4.9	256	65.6
RDCD(ours)	EFF-B0	4.7	256	65.7	link
RDCD(ours)	FastViT-T12	6.8	256	67.4

Requirements

Installation

pip install -r requirements.txt

Teacher Models(Pretrained)

• SSCD: ResNet-50, ResNeXt101
• DINO: ViT-B/8

Student Models

• EfficientNet-B0
• MobileNet-V3-Large
• FastViT-T12

How to use

docker install

docker build -t rdcd .
docker run --name rdcd -it --ipc=host --runtime=nvidia rdcd

Library install

pip install -c pytorch faiss-gpu
pip install Pillow==9.5.0

Training

sh train.sh

Evaluation

sh disc_eval.sh

Acknowledgement

Codebase from "SSCD" , pytorch-image-models

Citation

If you find this repository useful, please consider giving a star ⭐ and please cite as:

@inproceedings{kim2025relational,
  title={Relational Self-supervised Distillation with Compact Descriptors for Image Copy Detection},
  author={Kim, Juntae and Woo, Sungwon and Nang, Jongho},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  year={2025}
}