This repository is the official implementation of our AAAI 2024 Paper Vision Transformer Off-the-Shelf: A Surprising Baseline for Few-Shot Class-Agnostic Counting
Our code has been tested on Python 3.8.18 and PyTorch 1.8.1+cu111. Please follow the official instructions to setup your environment. See other required packages in requirements.txt.
We train and evaluate our methods on FSC-147 dataset. Please follow the FSC-147 official repository to download and unzip the dataset.
We have included our pretrained model here link. Then you can run following command to conduct the inference on the FSC-147 dataset.
python test.py
We use the same pretrained MAE model as CounTR, please download the pretrained MAE weight as CounTR. Then you can run the following command to conduct the traininng on the FSC-147 dataset.
python train_val.py
If you find this work or code useful for your research, please cite:
@inproceedings{wang2024vision,
title={Vision transformer off-the-shelf: A surprising baseline for few-shot class-agnostic counting},
author={Wang, Zhicheng and Xiao, Liwen and Cao, Zhiguo and Lu, Hao},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
year={2024}
}