Awesome
VSC-DescriptorTrack-Submission ($2^{nd}$ Solution)
The codes and related files to reproduce the results for Video Similarity Challenge Descriptor Track (CVPR2023).
Required dependencies
To begin with, you should install the packages according to the environment.yaml
file in this directory. Then install the GNU parallel
by sudo apt-get install parallel
(For Ubuntu). The minimum requirment for training is 8 Nvidia A100 40G GPUs, and for reference, you should have 1 Nvidia V100 16G GPU at least.
Pre-trained models
We use $7$ ImageNet-pre-trained models. Please download them from the provided links as below:
-
CotNet50: download_from_original_repo; The original project is CoTNet.
-
ResNet50: No need to download manually;
-
SKNet50: Google Drive; The original project is SKNet-PyTorch. Because it does not include the pre-trained models, please download the
sknet.py
file in this repository and follow the instruction in thePretrain
folder. -
Resnext50_32x4d: No need to download manually;
-
ViT: No need to download manually;
-
Swin: No need to download manually;
-
T2T: download_from_original_repo; The original project is T2T-ViT.
Training
For training, we propose Feature-Compatible Progressive Learning (FCPL). Please refer to the Training
folder for more details.
Test
The Test
folder shows the submitted code to extract query features plus the code to extract reference features and normalization features.
Citation
@article{wang2023feature,
title={Feature-compatible Progressive Learning for Video Copy Detection},
author={Wang, Wenhao and Sun, Yifan and Yang, Yi},
journal={arXiv preprint arXiv:2304.10305},
year={2023}
}
Bug finding
Please raise an issue or send an email to wangwenhao0716@gmail.com if a bug exists. Thanks!