InteractionCutMix

Interaction-aware Scene Debiasing Method for Action Recognition

1. Preparation

1. Clone this repository.

git clone --recursive https://github.com/rendicahya/intercutmix.git
cd intercutmix/

2. Create virtual environment.

python3 -m venv ~/venv/intercutmix/
source ~/venv/intercutmix/bin/activate
pip install -U pip

3. Install MMAction2.

pip install openmim
mim install mmengine mmcv
pip install -v -e mmaction2/

2. Download datasets

Link the project's data/ directory with MMAction2's data/ directory.

cd mmaction2/
ln -s ../data/ ./
cd -

a. UCF101

1. Download videos.

cd mmaction2/tools/data/ucf101/
bash download_videos.sh

This downloads UCF101.rar and unrars it to data/ucf101/videos.

2. Verify the number of videos.

find videos/ -type f | wc -l

Expected: 13,320.

3. Download annotations.

bash download_annotations.sh
cd -

This creates the annotations/ directory and puts files in it.

4. Generate splits.

python3 mmaction2/tools/data/build_file_list.py ucf101 data/ucf101/videos/ --format videos --shuffle --seed 0

This creates:

• ucf101_train_split_{1-3}_videos.txt
• ucf101_val_split_{1-3}_videos.txt

5. Verify files and directories.

intercutmix/data/ucf101/
├── annotations/
│   ├── classInd.txt
│   ├── testlist01.txt
│   ├── testlist02.txt
│   ├── testlist03.txt
│   ├── trainlist01.txt
│   ├── trainlist02.txt
│   └── trainlist03.txt
├── videos/
│   ├── ApplyEyeMakeup/
│   │   ├── v_ApplyEyeMakeup_g01_c01.avi
│   │   ├── v_ApplyEyeMakeup_g01_c02.avi
│   │   ├── v_ApplyEyeMakeup_g01_c03.avi
│   │   └── ...
│   └── ...
├── ucf101_train_split_1_videos.txt
├── ucf101_train_split_2_videos.txt
├── ucf101_train_split_3_videos.txt
├── ucf101_val_split_1_videos.txt
├── ucf101_val_split_2_videos.txt
└── ucf101_val_split_3_videos.txt

b. HMDB51

1. Download videos.

cd mmaction2/tools/data/hmdb51/
bash download_videos.sh

This downloads hmdb51_org.rar and unrars it to data/hmdb51/videos.

2. Verify the number of videos.

find videos/ -type f | wc -l

Expected: 6,766.

3. Download annotations.

bash download_annotations.sh
cd -

This creates the annotations/ directory and puts files in it.

4. Generate splits.

python3 mmaction2/tools/data/build_file_list.py hmdb51 data/hmdb51/videos/ --format videos --shuffle --seed 0

This creates:

• hmdb51_train_split_{1-3}_videos.txt
• hmdb51_val_split_{1-3}_videos.txt

5. Verify files and directories.

intercutmix/data/hmdb51/
├── annotations/
│   ├── brush_hair_test_split1.txt
│   ├── brush_hair_test_split2.txt
│   ├── brush_hair_test_split3.txt
│   └── ...
├── frames/
│   ├── brush_hair/
│   │   ├── April_09_brush_hair_u_nm_np1_ba_goo_0/
│   │   │   ├── img_00001.png
│   │   │   ├── img_00002.png
│   │   │   ├── img_00003.png
│   │   │   └── ...
│   │   └── ...
│   └── ...
├─── videos/
│   ├── brush_hair/
│   │   ├── April_09_brush_hair_u_nm_np1_ba_goo_0.avi
│   │   ├── April_09_brush_hair_u_nm_np1_ba_goo_1.avi
│   │   ├── April_09_brush_hair_u_nm_np1_ba_goo_2.avi
│   │   └── ...
│   └── ...
├── hmdb51_train_split_1_videos.txt
├── hmdb51_train_split_2_videos.txt
├── hmdb51_train_split_3_videos.txt
├── hmdb51_val_split_1_videos.txt
├── hmdb51_val_split_2_videos.txt
└── hmdb51_val_split_3_videos.txt

c. Kinetics100

1. Download videos.

mkdir -p data/kinetics100
cd data/kinetics100
gdown 1_gPSZDo_yasyEbtfs0m2edZAzhsr8dU4
tar xzf kinetics100.tar.gz
mv kinetics_100/ videos/
rm kinetics100.tar.gz

This downloads kinetics100.tar.gz and extracts it to data/kinetics100/videos.

2. Verify the number of videos.

find videos/ -type f | wc -l

Expected: 9,999.

3. Open settings.json and set active.dataset: kinetics100.

4. Generate annotations.

python3 tools/data/classind.py

This creates data/kinetics100/annotations/classInd.txt.

5. Generate splits.

python3 tools/data/split.py

This creates:

• data/kinetics100/kinetics100_train_split_1_videos.txt
• data/kinetics100/kinetics100_val_split_1_videos.txt

3. Inference

1. Download checkpoints.

pip install gdown
mkdir checkpoints/
gdown -O checkpoints/ <download-key>

Refer the following table for <download-key>.

2. Run inference.

python3 mmaction/tools/test.py <config-path> <checkpoint-path>

Refer the following table for <config-path> and <checkpoint-path>.

Dataset	Top-1	Top-5	Config Path	Checkpoint Path	Download Key
UCF101	87.84%	97.12%	mmaction2/configs/recognition/c3d-ucf101-soft/c3d_sports1m-pretrained_8xb64-16x1x1-100e_ucf101-rgb-intercutmix-p0.5-mmr0.05-a2.py	checkpoints/ucf101-icm-p0.5-mmr0.05-a2.pth	1Aynmc64VpLJEXBeNe-Bq_u787h5pX2aq
HMDB51	55.75%	85.49%	mmaction2/configs/recognition/c3d-hmdb51-soft/c3d_sports1m-pretrained_8xb64-16x1x1-100e_hmdb51-rgb-intercutmix-p0.5-mmr0.05-a2.py	checkpoints/hmdb51-icm-p0.5-mmr0.05.pth	1cb3gfG3qJUAAsrMSXcXm0LKHLSuuDGd6
Kinetics100

Cite

@ARTICLE{11045925,
  author={Wihandika, Randy Cahya and Mendonça, Israel and Aritsugi, Masayoshi},
  journal={IEEE Access}, 
  title={Interaction-Aware Scene Debiasing for Action Recognition}, 
  year={2025},
  volume={13},
  number={},
  pages={107856-107871},
  keywords={Semantics;Training;Punching;Data models;Data augmentation;Visualization;Correlation;Training data;Spatiotemporal phenomena;Representation learning;Action recognition;scene debiasing;video augmentation},
  doi={10.1109/ACCESS.2025.3581931}}