OpenThermalAction: An Open-Source Dataset of Thermal Human Action Videos with Frame-Level Skeleton Annotations

The dataset comprises 2,488 thermal videos collected from 55 subjects performing 28 action classes, including both single-person sports activities and multi-person daily interactions.

Structure

├── train/
│   ├── session_1/            # Session 1: Single-person activities
│   │   ├── sub_1/
│   │   │   └── thermal/
│   │   │       ├── 1_1_0/    # Action clips as image folders
│   │   │       ├── 1_2_0/
│   │   │       └── ...
│   │   ├── sub_2/
│   │   └── ...
│   └── session_2/            # Session 2: Two-person interactions
│       ├── exp_1/
│       │   └── thermal/
│       │       ├── 1/
│       │       ├── 2/
│       │       └── ...
│       └── ...
├── val/
│   ├── session_1/
│   └── session_2/
├── test/
│   ├── test_s1/              # Session 1: Single-person activities
│   │   ├── sub_4/
│   │   │   └── thermal/
│   │   │       ├── 1_1/      # Action clips as image folders
│   │   │       ├── 1_2/
│   │   │       └── ...
│   │   ├── sub_7/
│   │   ├── sub_14/
│   │   └── ...
│   └── test_s2/              # Session 2: Two-person interactions
│       ├── exp_7/
│       │   └── thermal/
│       │       ├── 1/
│       │       ├── 2/
│       │       └── ...
│       └── ...
├── test_wild/
│   ├── test_s1/              # Session 1: Single-person activities
│   │   ├── sub_1/
│   │   │   └── thermal/
│   │   │       ├── 1_1/      # Action clips as image folders
│   │   │       ├── 1_2/
│   │   │       └── ...
│   │   ├── sub_2/
│   │   ├── sub_3/
│   │   └── ...
│   └── test_s2/              # Session 2: Two-person interactions
│       ├── exp_1/
│       │   └── thermal/
│       │       ├── 1/
│       │       ├── 2/
│       │       └── ...
│       └── ...
├── pyskl_annotations
├── annotations_train.txt
├── annotations_val.txt
├── annotations_test.txt
└── annotations_wild_test.txt

Archived Files

For easier download, each subject/experiment is archived separately:

Session 1 (Single-Person Actions):

• sub_1.tar.gz
• sub_2.tar.gz
• sub_3.tar.gz
• ... (one archive per subject)

Session 2 (Multi-Person Actions):

• exp_1.tar.gz
• exp_2.tar.gz
• ... (one archive per experiment)

Annotation Format

Each line in annotations_train.txt, annotations_val.txt, and annotations_test.txt contains:

path_to_sequence action_class

Example:

test_s1/sub_4/thermal/1_1 0

action_class (0-14): Sports activities

action_class (15-27): Daily activities

Skeleton Annotations in the PYSKL-format

The dataset was annotated with bounding boxes and 17-keypoints in the COCO format. We arranged skeletons in the PYSKL format to train using the PYSKL library.

Structure

pyskl_annotations/
├── pyskl_ota.pkl (full dataset)
├── pyskl_ota_75.pkl (75% of frames)
├── pyskl_ota_50.pkl (50% of frames)
├── pyskl_ota_25.pkl (25% of frames)
├── pyskl_ota_session1.pkl (full dataset, session 1 only)
├── pyskl_ota_session2.pkl (full dataset, session 2 only)
├── pyskl_ota_75_session1.pkl (75% frames, session 1)
├── pyskl_ota_75_session2.pkl (75% frames, session 2)
├── pyskl_ota_50_session1.pkl (50% frames, session 1)
├── pyskl_ota_50_session2.pkl (50% frames, session 2)
├── pyskl_ota_25_session1.pkl (25% frames, session 1)
├── pyskl_ota_25_session2.pkl (25% frames, session 2)
├── ...

Full Dataset

• pyskl_ota.pkl: Full dataset with all frames (train/val from Balcony, test from Office)
• pyskl_ota_session1.pkl: Full dataset, session 1 only (train/val: session_1, test: test_s1)
• pyskl_ota_session2.pkl: Full dataset, session 2 only (train/val: session_2, test: test_s2)

Frame-Reduced (75%)

• pyskl_ota_75.pkl: Frame-reduced to first 75% of frames
• pyskl_ota_75_session1.pkl: 75% frames, session 1 only
• pyskl_ota_75_session2.pkl: 75% frames, session 2 only

Frame-Reduced (50%)

• pyskl_ota_50.pkl: Frame-reduced to first 50% of frames
• pyskl_ota_50_session1.pkl: 50% frames, session 1 only
• pyskl_ota_50_session2.pkl: 50% frames, session 2 only

Frame-Reduced (25%)

• pyskl_ota_25.pkl: Frame-reduced to first 25% of frames
• pyskl_ota_25_session1.pkl: 25% frames, session 1 only
• pyskl_ota_25_session2.pkl: 25% frames, session 2 only

Github link: https://github.com/IS2AI/OpenThermalAction

Models link: https://huggingface.co/issai/thermal-skeleton-models