This repo holds the codes and models for the AME framework presented on ACM MM2023
Exploring Motion Cues for Video Test-Time Adaptation Runhao Zeng, Qi Deng, Huixuan Xu, Shuaicheng Niu, Jian Chen, ACM MM ’23, October 29-November 3, 2023, Ottawa, ON, Canada
The training and testing in AME is reimplemented in PyTorch for the ease of use.
- PyTorch Other minor Python modules can be installed by running
pip install -r requirements.txt
git clone --recursive https://github.com/Alvin-Zeng/AME
You can use this command to training and testing AME
You can use a for loop iterates to control which noise you want to train and test AME.If you want to change other noise,you can replace contrast with the name of another noise.
for CORRUPT in contrast
do
CUDA_VISIBLE_DEVICES=7 python main.py \
--seed=507 \
--log_dir=log_seed/tanet_ucf101/${CORRUPT}/5e-6 \
--time_log \
--dataset=ucf101-${CORRUPT} \
--checkpoint=$PATH_OF_TRAINING_CHECKPOINT \
--dataset_path= $PATH_OF_TRAINING_DATASET \
--save_ckpt= $PATH_OF_SAVING_TRAINING_CHECKPOINT \
--mix \
--lr=5e-6 \
--gpus 0
done
You can also use other commands in the script folder to train different dataset
Comparisons of test-time adaptation performance on UCF101 dataset. * video domain adaptation method.
Method | gauss | pepper | salt | shot | zoom | impulse | motion | jpeg | contrast | rain | h265.abr | avg |
---|---|---|---|---|---|---|---|---|---|---|---|---|
Without Adaptation | 17.50 | 23.05 | 6.85 | 71.82 | 75.55 | 16.94 | 54.77 | 82.92 | 62.89 | 81.31 | 78.54 | 51.98 |
BN Adaptation | 37.01 | 33.49 | 20.64 | 80.01 | 76.13 | 37.59 | 54.46 | 83.08 | 69.13 | 85.85 | 76.90 | 59.57 |
NORM | 41.79 | 39.70 | 22.26 | 84.54 | 80.63 | 43.38 | 61.55 | 88.00 | 70.82 | 89.29 | 80.97 | 63.90 |
Contrast TTA | 36.58 | 27.57 | 21.33 | 74.31 | 69.79 | 36.11 | 49.48 | 80.23 | 24.48 | 78.46 | 74.60 | 52.09 |
SAR | 48.48 | 43.00 | 22.60 | 85.30 | 68.60 | 35.40 | 40.43 | 86.41 | 64.93 | 81.55 | 77.39 | 59.46 |
ATCoN* | 60.19 | 50.60 | 32.60 | 84.80 | 78.80 | 62.50 | 69.40 | 84.70 | 71.10 | 86.30 | 78.30 | 69.03 |
Ours | 72.06 | 64.45 | 53.50 | 86.84 | 77.80 | 67.09 | 63.57 | 88.94 | 71.76 | 90.50 | 80.89 | 74.31 |
Please cite the following paper if you feel AME useful to your research
@inproceedings{ACMMM-AME,
author = {Runhao Zeng and
Qi Deng and
Huixuan Xu and
Shuaicheng Niu and
Jian Chen},
title = {Exploring Motion Cues for Video Test-Time Adaptation},
booktitle = {ACM MM2023},
year = {2023},
}
For any question, please file an issue or contact
Runhao Zeng: runhaozeng.cs@gmail.com