Awesome
CounTR
Official PyTorch implementation for CounTR. Details can be found in the paper. [Paper] [Project page]
Thanks to @tamnguyenvan, now we can use CounTR in an easier way by using this library!
<img src=img/arch.png width="80%"/>
Contents
- Preparation
- CounTR train
- CounTR inference
- Fine-tuned weights
- Visualisation
- Citation
- Acknowledgements
Preparation
1. Download datasets
In our project, the following datasets are used. Please visit following links to download datasets:
In fact, we use CARPK by importing hub package. Please click here for more information.
2. Download required python packages:
The following packages are suitable for NVIDIA GeForce RTX 3090.
pip install torch==1.10.0+cu111 torchvision==0.11.0+cu111 torchaudio==0.10.0 -f https://download.pytorch.org/whl/torch_stable.html
pip install timm==0.3.2
pip install numpy
pip install matplotlib tqdm
pip install tensorboard
pip install scipy
pip install imgaug
pip install opencv-python
pip3 install hub
- This repo is based on
timm==0.3.2
, for which a fix is needed to work with PyTorch 1.8.1+.
CounTR Train
Please modify your work directory and dataset directory in the following train files.
Task | model file | train file |
---|---|---|
Pretrain on FSC147 | models_mae_noct.py | FSC_pretrain.py |
Finetune on FSC147 | models_mae_cross.py | FSC_finetune_cross.py |
Finetune on CARPK | models_mae_cross.py | FSC_finetune_CARPK.py |
Pretrain on FSC147
CUDA_VISIBLE_DEVICES=0 python FSC_pretrain.py \
--epochs 500 \
--warmup_epochs 10 \
--blr 1.5e-4 --weight_decay 0.05
Finetune on FSC147
CUDA_VISIBLE_DEVICES=0 nohup python -u FSC_finetune_cross.py \
--epochs 1000 \
--blr 2e-4 --weight_decay 0.05 >>./train.log 2>&1 &
Finetune on CARPK
CUDA_VISIBLE_DEVICES=0 nohup python -u FSC_finetune_CARPK.py \
--epochs 1000 \
--blr 2e-4 --weight_decay 0.05 >>./train.log 2>&1 &
CounTR Inference
Please modify your work directory and dataset directory in the following test files.
Task | model file | test file |
---|---|---|
Test on FSC147 | models_mae_cross.py | FSC_test_cross.py |
Test on CARPK | models_mae_cross.py | FSC_test_CARPK.py |
Test on FSC147
CUDA_VISIBLE_DEVICES=0 nohup python -u FSC_test_cross.py >>./test.log 2>&1 &
Test on CARPK
CUDA_VISIBLE_DEVICES=0 nohup python -u FSC_test_CARPK.py >>./test.log 2>&1 &
Also, demo.py is a small demo used for testing on a single image.
CUDA_VISIBLE_DEVICES=0 python demo.py
Fine-tuned weights
benchmark | MAE | RMSE | link |
---|---|---|---|
FSC147 | 11.95 (Test set) | 91.23 (Test set) | weights |
CARPK | 5.75 | 7.45 | weights |
Visualisation
<img src=img/goodpred.png width="75%"/>
Citation
@inproceedings{liu2022countr,
author = {Chang, Liu and Yujie, Zhong and Andrew, Zisserman and Weidi, Xie},
title = {CounTR: Transformer-based Generalised Visual Counting},
booktitle={British Machine Vision Conference (BMVC)},
year = {2022}
}
Acknowledgements
We borrowed the code from
Thanks @GioFic95 for adding the function of using external exemplars, more predictions images, more parametrized inference and so on.
If you have any questions about our code implementation, please contact us at liuchang666@sjtu.edu.cn