🧼 SOAP (Under Construction)

Code for "SOAP: Enhancing Spatio-Temporal Relation and Motion Information Capturing for Few-Shot Action Recognition", accepted by ACM MM'24.
Paper

📹 Usage

Requirements -> Dataset Preparation -> Training -> Testing

Requirements

python >= 3.6
pytorch >= 1.8

Data preparation

SthSthV2
Kinetics
UCF101

HMDB51

The data splts .txt files

splits
|_ hmdb_ARN
| |_ test.txt
| | |_ [class0]/[video0]
| | |_ ...
| | |_ [class1]/[video0]
| | |_ ...
| |_ train.txt
| |_ val.txt
|_ kinetics_CMN
|_ ssv2_OTAM
|_ ucf_ARN

The pre-processed dataset is organized with the following structure:

  datasets
   |_ Kinetics
   | |_ frames
   |   |_ [class_0]
   |   | |_ [video_0]
   |   | | |_ img_00001.jpg
   |   | | |_ img_00001.jpg
   |   | | |_ ...
   |   | |_ [video_1]
   |   | | |_ img_00001.jpg
   |   | | |_ img_00002.jpg
   |   | | |_ ...
   |   |_ [class_1]
   |   |_ ...
   |_ ...

⚙ Training Example

run_list.sh

python runfull.py --way 5 --shot 1 --dataset kinetics --training_iterations 10000 --learning_rate 0.001 --sch 2000 6000 8000
python runfull.py --way 5 --shot 5 --dataset kinetics --training_iterations 10000 --learning_rate 0.001 --sch 2000 6000 8000
python runfull.py --way 5 --shot 1 --dataset hmdb --training_iterations 10000 --learning_rate 0.001 --sch 2000 6000 8000
python runfull.py --way 5 --shot 5 --dataset hmdb --training_iterations 10000 --learning_rate 0.001 --sch 2000 6000 8000
python runfull.py --way 5 --shot 1 --dataset ucf --training_iterations 10000 --learning_rate 0.001 --sch 2000 6000 8000
python runfull.py --way 5 --shot 5 --dataset ucf --training_iterations 10000 --learning_rate 0.001 --sch 2000 6000 8000
python runfull.py --way 5 --shot 1 --dataset ssv2 --training_iterations 75000 --learning_rate 0.001 --sch 20000 40000 60000
python runfull.py --way 5 --shot 5 --dataset ssv2 --training_iterations 75000 --learning_rate 0.001 --sch 20000 40000 60000

🚀 Testing Example

test_list.sh

python testfull.py --way 5 --shot 1 --dataset kinetics --best_model_path 'best_model/kinetics/1-shot/checkpoint_best_val.pt' > ./test_log/kinetics_1shot.txt
python testfull.py --way 5 --shot 5 --dataset kinetics --best_model_path 'best_model/kinetics/5-shot/checkpoint_best_val.pt' > ./test_log/kinetics_5shot.txt
python testfull.py --way 5 --shot 1 --dataset ssv2 --best_model_path 'best_model/ssv2/1-shot/checkpoint_best_val.pt' > ./test_log/ssv2_1shot.txt
python testfull.py --way 5 --shot 5 --dataset ssv2 --best_model_path 'best_model/ssv2/5-shot/checkpoint_best_val.pt' > ./test_log/ssv2_5shot.txt
python testfull.py --way 5 --shot 1 --dataset hmdb --best_model_path 'best_model/hmdb/1-shot/checkpoint_best_val.pt' > ./test_log/hmdb_1shot.txt
python testfull.py --way 5 --shot 5 --dataset hmdb --best_model_path 'best_model/hmdb/5-shot/checkpoint_best_val.pt' > ./test_log/hmdb_5shot.txt
python testfull.py --way 5 --shot 1 --dataset ucf --best_model_path 'best_model/ucf/1-shot/checkpoint_best_val.pt' > ./test_log/ucf_1shot.txt
python testfull.py --way 5 --shot 5 --dataset ucf --best_model_path 'best_model/ucf/5-shot/checkpoint_best_val.pt' > ./test_log/ucf_5shot.txt

⚠️ Copyright

People from my previous group (from 2022.4 to 2023.9). If you open this repo, thank to me (Wenbo Huang, code holder) by default.
来自我前课题组（2022.4至2023.9）的人员。倘若打开这一仓库，默认对我(黄文博，代码持有者)表示感谢。

wenbohuang1002 / SOAP

readme