Delete README.md
Browse files
README.md
DELETED
|
@@ -1,92 +0,0 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: apache-2.0
|
| 3 |
-
---
|
| 4 |
-
|
| 5 |
-
|
| 6 |
-
# Dataset Card for ASTime (Activity State Transition Timestamp)
|
| 7 |
-
|
| 8 |
-

|
| 9 |
-
|
| 10 |
-
The Activity State Transition Timestamp (ASTime) dataset is constructed from a robotic perspective for real-time perception of human activity changes. Unlike traditional datasets that only mark the start and end points of an activity, ASTime introduces a novel annotation scheme that explicitly labels the activity state transition timestamp, defined as the exact moment when a new activity becomes semantically distinguishable from the preceding one.
|
| 11 |
-
|
| 12 |
-
## Dataset Details
|
| 13 |
-
|
| 14 |
-
### Dataset Description
|
| 15 |
-
|
| 16 |
-
ASTime comprises 161 untrimmed long videos totaling 14 hours and 37 minutes, recorded at a resolution of 1920x1080 and 30 FPS, containing approximately 1.72 million frames. From these continuous streams, we annotated 3,469 continuous video samples (2,924 for training and 545 for testing).
|
| 17 |
-
|
| 18 |
-
The dataset is collected from 16 human subjects exhibiting diverse behavioral patterns and visual appearances. All videos are recorded with unrestricted viewpoints and camera placements across multiple simulated home scenarios such as living rooms, bedrooms, and kitchens.
|
| 19 |
-
|
| 20 |
-
It covers 44 activity types, combining 18 common activities (e.g., reading, falling, headache) with three basic postures: standing, sitting, and lying.
|
| 21 |
-
|
| 22 |
-
<!-- - **Curated by:** [Insert Author/Lab Name]
|
| 23 |
-
- **Funded by [optional]:** [Insert Funding Agency/Grant]
|
| 24 |
-
- **Shared by [optional]:** [Insert Affiliation] -->
|
| 25 |
-
- **Language(s) (NLP):** English (Annotations)
|
| 26 |
-
- **License:** Apache 2.0
|
| 27 |
-
|
| 28 |
-
<!-- ### Dataset Sources [optional]
|
| 29 |
-
|
| 30 |
-
<!-- - **Repository:** [Insert GitHub/HuggingFace Link] -->
|
| 31 |
-
<!-- - **Paper [optional]:** [https://github.com/Jambo-12/ActPro] -->
|
| 32 |
-
<!-- - **Demo [optional]:** [Insert Demo Link if available] --> -->
|
| 33 |
-
|
| 34 |
-
## Uses
|
| 35 |
-
|
| 36 |
-
### Direct Use
|
| 37 |
-
|
| 38 |
-
This dataset is primarily designed for:
|
| 39 |
-
- **Activity State Transition Detection:** Identifying the exact timestamp when a semantic shift in human activity occurs.
|
| 40 |
-
- **Real-time Robotic Perception:** Assisting robotic systems in monitoring and reacting to continuous human activities in home environments.
|
| 41 |
-
<!-- - **Elder Care Monitoring:** Detecting safety-critical anomalous events like sudden illness (e.g., stomachache, headache) or falls. -->
|
| 42 |
-
- **Dense Video Captioning & Action Recognition:** Utilizing the previous activity's semantic caption (`t1_caption`) as historical context to predict subsequent frames and actions.
|
| 43 |
-
|
| 44 |
-
### Out-of-Scope Use
|
| 45 |
-
|
| 46 |
-
- **Socio-economic Diversity Studies:** The dataset features demographic homogeneity (primarily graduate students) and is not suited for studying broad socio-economic differences.
|
| 47 |
-
- **Biometric Identification/Surveillance:** Facial information is intended solely for algorithmic analysis and should not be linked to participant identities for surveillance purposes.
|
| 48 |
-
- **Modeling Naturalistic Frequency:** The dataset employs an "Intentional Over-sampling Strategy" for rare/critical events. Therefore, it does not reflect the natural, real-world long-tailed frequency of daily activities.
|
| 49 |
-
|
| 50 |
-
## Dataset Structure
|
| 51 |
-
|
| 52 |
-
### Video Data
|
| 53 |
-
- **Total Scale:** 161 untrimmed long-form videos.
|
| 54 |
-
- **Duration:** 14 hours and 37 minutes total.
|
| 55 |
-
- **Frame Count:** ~1.72 million frames.
|
| 56 |
-
- **Resolution/Framerate:** 1920x1080 at 30 FPS.
|
| 57 |
-
|
| 58 |
-
### Annotation Format
|
| 59 |
-
Annotations are provided in JSON format, structured by video filename. Example:
|
| 60 |
-
```json
|
| 61 |
-
{
|
| 62 |
-
"lsr1-1": [
|
| 63 |
-
{
|
| 64 |
-
"t1_caption": "standing, pouring",
|
| 65 |
-
"start_time": 0.5,
|
| 66 |
-
"timestamp": 3.0,
|
| 67 |
-
"end_time": 12.5,
|
| 68 |
-
"text": "sitting, using a laptop",
|
| 69 |
-
"flag": 0
|
| 70 |
-
}
|
| 71 |
-
]
|
| 72 |
-
}
|
| 73 |
-
```
|
| 74 |
-
|
| 75 |
-
### Training-Ready Annotations (Captioned by Qwen2.5VL)
|
| 76 |
-
|
| 77 |
-
In addition to the standard categorical labels, we also provide a training-ready annotation format. In this version, the activity labels (`text`) and the historical context (`t1_caption`) have been enriched into detailed natural language descriptive sentences using **Qwen2.5VL**. This format is particularly well-suited for training Large Language Models (LLMs) or Vision-Language Models (VLMs) for video understanding tasks.
|
| 78 |
-
|
| 79 |
-
**Example Format:**
|
| 80 |
-
```json
|
| 81 |
-
{
|
| 82 |
-
"wwl1-1": [
|
| 83 |
-
{
|
| 84 |
-
"t1_caption": "The person appears to be lying on the couch, possibly experiencing a headache.",
|
| 85 |
-
"start_time": 38.79,
|
| 86 |
-
"timestamp": 49.42,
|
| 87 |
-
"end_time": 56.30,
|
| 88 |
-
"text": "The person is standing and drinking from a cup.",
|
| 89 |
-
"flag": 0
|
| 90 |
-
}
|
| 91 |
-
]
|
| 92 |
-
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|