README.md

---
license: cc-by-nc-4.0
task_categories:
  - other
tags:
  - gesture-recognition
  - action-recognition
  - multimodal
  - IMU
  - video
size_categories:
  - 100K<n<1M
---

# Directory Structure

Published data is organized by collection batch ID.

```text
HiSync_publish/
├── 1/ # Batch ID
│   ├── user1_20250726_132447/ # Sample directory
│   │   ├── cam_1/
│   │   ├── cam_2/
│   │   ├── cam_3/
│   │   ├── person_keypoints.json
│   │   └── meta.json
│   ├── user2_20250726_135244/
│   │   └── ...
│   └── IMU/
│       ├── IMU_Palm/
│       │   └── *.csv
│       ├── IMU_Ring/
│       │   └── *.csv
│       └── IMU_Wrist/
│           └── *.csv
├── 2/
│   └── ...
└── 18/
    └── ...
```

Notes:

1. Each sample directory is named `userX_YYYYMMDD_HHMMSS`.
2. Each sample directory contains camera data, `person_keypoints.json`, and `meta.json`.
3. IMU data is aggregated per batch under `batch_id/IMU/IMU_{Palm|Ring|Wrist}`, not in individual sample directories.

# Data Format

Example `meta.json` for a sample:

```json
{
	"user": "user10",
	"action": "Right",
	"perspective": "Eye-level",
	"distance": "10-15m",
	"camera": {
        "cam_0": "telephone",
        "cam_2": "iphone",
        "cam_1": "cam"
    },
	"IMU": {
		"Palm": {
            "timestamp": "5/IMU/IMU_Palm/calibrated_imu_20250727_150254.csv"
        },
        "Ring": {
            "timestamp": "5/IMU/IMU_Ring/calibrated_imu_20250727_150254.csv"
        },
        "Wrist": {
            "timestamp": "5/IMU/IMU_Wrist/calibrated_imu_20250727_150254.csv"
        }
	}
}
```

Field Constraints:

1. `action` is standardized to: `Right`, `Left`, `Approach`, `Retreat`, `Summon`, `Ascend`, `Descend`, `No-Gesture`.
2. `perspective` is standardized to: `Upward`, `Eye-level`, `Downward`.
3. `distance` is standardized as range strings: `3-5m`, `5-10m`, `10-15m`, `15-20m`, `20-25m`, `25-34m`.
4. `IMU.*.timestamp` may be `null`; handle null values during parsing.

> A small portion of the data may have missing camera or IMU modalities. Ensure robust error handling during reading.


Example `person_keypoints.json`:

```json
{
    "0": [
        {
            "frame_idx": 0,
            "filename": "frame_0000.png",
            "keypoints": [
                [1204.31, 181.52, 0.9949],
                [1208.77, 170.23, 0.9781],
                [1198.12, 170.84, 0.9675],
                [1216.43, 181.65, 0.8360],
                [1187.55, 183.20, 0.6951]
            ],
            "bbox": [1133, 106, 1373, 813]
        },
        {
            "frame_idx": 1,
            "filename": "frame_0001.png",
            "keypoints": [
                [1204.88, 181.46, 0.9955],
                [1209.06, 170.34, 0.9761],
                [1197.93, 170.95, 0.9748]
            ],
            "bbox": [1132, 106, 1374, 813]
        }
    ],
    "1": [
        {
            "frame_idx": 0,
            "filename": "frame_0000.png",
            "keypoints": [],
            "bbox": [0, 0, 0, 0]
        }
    ]
}
```

Notes:

1. Top-level keys (e.g., `"0"`, `"1"`) represent person IDs (string format).
2. Each camera corresponds to a frame list with elements containing `frame_idx`, `filename`, `keypoints`, and `bbox`.
3. Each point in `keypoints` is `[x, y, score]`, following COCO-17 order.
4. When no person is detected in a frame, `keypoints` may be an empty array; handle this during parsing.