ASTime

ASTime, short for Activity State Transition Timestamp, is a video dataset for studying real-time perception of human activity changes from a robotic perspective. Instead of only providing the start and end time of an activity segment, ASTime explicitly annotates the transition timestamp: the moment at which a new human activity state becomes semantically distinguishable from the previous state.

The released dataset contains anonymized, muted, untrimmed long videos collected in constructed home-like laboratory environments, together with JSON annotations for activity-state transition timestamp localization. The train/test split is subject-disjoint to evaluate generalization to unseen participants.

📌 Dataset Summary

The statistics below are computed from the files in this repository under Videos/ and Annotations/.

Split	Subjects	Videos	Standard annotation JSON files	Transition events	Duration	Frames
Train	14	128	128	2,924	13:20:56.116	1,440,358
Test	2	33	33	545	02:37:54.942	283,964
Total	16	161	161	3,469	15:58:51.058	1,724,322

Quick file facts:

• 🎞️ Video format: MP4 / H.264.
• 🔇 Audio: no audio streams are present in the released videos.
• 🖼️ Resolution and frame rate: most videos are 1920 x 1080 at about 30 FPS; two videos are 3840 x 2160 at 30 FPS.

Label vocabulary:

• 3 basic postures: standing, sitting, and lying.
• 18 non-posture action terms, such as reading, using a phone, falling, stomachache, and carrying heavy objects.
• Composite labels combine a posture with one or more action terms, for example sitting, using a phone or sitting, using a phone, eating.

🗂️ Dataset Structure

The dataset has two main components: video files and JSON annotations.

Component	Path pattern	Description
🎬 Training videos	Videos/train/P001_train/*.MP4 ... Videos/train/P014_train/*.MP4	Untrimmed videos for the training subjects.
🎬 Test videos	Videos/test/P015_test/*.MP4, Videos/test/P016_test/*.MP4	Untrimmed videos for the held-out test subjects.
🏷️ Training labels	Annotations/ASTime_Annotations_train/annotation_train_labels/*.json	Standard activity-state transition labels for training videos.
🏷️ Test labels	Annotations/ASTime_Annotations_test/annotation_test_label/*.json	Standard activity-state transition labels for test videos.
💬 Training captions	Annotations/ASTime_Annotations_train/annotation_train_captioned/*.json	Auxiliary natural-language captioned annotations for training videos.

The data/ directory contains a normalized Parquet representation of the training caption annotations for Hugging Face Dataset Viewer visualization. The original videos and JSON annotation files remain the canonical source files under Videos/ and Annotations/, respectively.

Naming follows the same anonymized subject/video ID convention across videos and annotations:

Item	Example
Subject folder	P001_train, P015_test
Video file	P001_train_1-1.MP4
Standard label file	P001_train_1-1.json
Video ID / JSON key	P001_train_1-1

The subject IDs are anonymized as P001 through P016. The training split contains P001 to P014, and the testing split contains P015 and P016.

📝 Annotation Format

Annotations are stored as JSON files. Each JSON file contains a single top-level key, which is the video ID. The value is a list of transition-event annotations.

Video-Annotation Matching

Use the shared video ID, i.e. the file stem without extension, to locate the corresponding standard annotation file.

Video	Annotation
Videos/train/P001_train/P001_train_1-1.MP4	Annotations/ASTime_Annotations_train/annotation_train_labels/P001_train_1-1.json
Videos/test/P015_test/P015_test_1-1.MP4	Annotations/ASTime_Annotations_test/annotation_test_label/P015_test_1-1.json

Standard Label Example

Standard label files use compact activity-state labels in the text field:

{
  "P001_train_1-1": [
    {
      "t1_caption": "",
      "start_time": 8.0,
      "timestamp": 11.0,
      "end_time": 19.0,
      "text": "sitting, using a laptop",
      "flag": 0
    }
  ]
}

Captioned Annotation Example

The training split also includes an auxiliary captioned version in annotation_train_captioned/. In these files, t1_caption describes the previous state and text describes the target state in natural language:

{
  "P001_train_1-1": [
    {
      "t1_caption": "This person is sitting.",
      "start_time": 8.0,
      "timestamp": 11.0,
      "end_time": 19.0,
      "text": "The person is sitting on a couch and using a laptop.",
      "flag": 0
    }
  ]
}

Field definitions:

Field	Type	Description
t1_caption	string	Previous-state natural-language caption when available. This field is empty in the standard label annotations and populated in the auxiliary training caption files.
start_time	float	Start time, in seconds, of the temporal window associated with the transition event.
timestamp	float	Activity-state transition timestamp, in seconds. This is the target timestamp to localize.
end_time	float	End time, in seconds, of the temporal window associated with the transition event.
text	string	Target activity-state label or natural-language target-state caption, depending on the annotation directory.
flag	integer	Auxiliary annotation flag retained from the annotation process. In the current release, 3,413 events have flag = 0 and 56 events have flag = 1.

For the main activity-state transition timestamp task, use:

• Training labels: Annotations/ASTime_Annotations_train/annotation_train_labels/
• Test labels: Annotations/ASTime_Annotations_test/annotation_test_label/

The directory Annotations/ASTime_Annotations_train/annotation_train_captioned/ provides an auxiliary natural-language captioned version of the training annotations. In those files, text is a natural-language target-state caption and t1_caption describes the previous state.

All standard label annotations satisfy:

start_time <= timestamp <= end_time

🛠️ Dataset Construction and Collection Protocol

Environment and Setup

We constructed a comprehensive simulated home environment within a laboratory to serve as the recording space. The setup comprises eight distinct functional rooms, including three living rooms, one bedroom, one kitchen, two study rooms, and one conference room. To introduce spatial variance, the internal furniture layout within each room was randomly adjusted across different recording sessions. The lighting conditions captured natural variations, encompassing both daytime and nighttime illumination. Regarding the recording equipment, we avoided static surveillance setups. For every specific scene, two cameras were dynamically deployed at different positions and viewing angles, specifically chosen to ensure clear capture of the human activities without adhering to a fixed coordinate system.

Participants and Execution

The dataset features 16 healthy recruited university students aged between 22 and 31, comprising 14 males and 2 females. All participants reported no physical or mobility impairments, ensuring baseline kinematic consistency across the recorded routines. Furthermore, they wore their own normal daily clothing to introduce natural visual variance, and rigorous decency standards were maintained throughout all sessions. Crucially, there were no rehearsals prior to the recordings to avoid rigid or performative movements. Rather than executing isolated and mechanized actions, participants were instructed to naturally transition between multiple randomly assigned activities, ensuring long and continuous untrimmed video sequences.

Mitigating Observer Bias via Blind Protocol

To ensure the ecological validity of the collected data and minimize potential observer effects, we employed a task-oriented blind protocol during the recording process:

• Blinded to Measurement Target: Participants were informed that the study aimed to evaluate general robotic perception capabilities rather than specific timing accuracy or transition timestamp detection. This minimized demand characteristics, preventing participants from intentionally exaggerating or pausing at activity boundaries.
• High-Level Directives: Participants were not given micro-level kinematic step-by-step scripts. Instead, they received high-level task directives. For example, rather than instructing a participant to walk to the sofa, sit down, wait three seconds, and then pick up the book, the blind directive simply stated go read a book in the living room for a while. Similarly, for safety-critical events, the instruction was simulate a sudden stomachache while cleaning rather than dictating specific falling mechanics. This approach preserved the natural diversity of movement execution, ensuring that postural transitions and intermediate hesitations reflected natural human behavioral dynamics.

Annotation Protocol and Boundary Disambiguation

To balance annotation efficiency and precision, we implemented a two-stage annotation protocol:

1. Real-Time Initial Localization: During data collection, a custom real-time logging tool recorded the preliminary start and end times of each activity sequence.
2. Multi-Annotator Fine-Grained Alignment: Three independent graduate researchers trained in video analysis conducted post-hoc refinement to ensure the objectivity of the novel transition timestamp.

When facing ambiguous boundaries (e.g., the transition between reaching for a book and reading), the annotators adhered to a strict semantic establishment criterion. The transition timestamp was defined as the exact moment when the new behavioral state was fully established and visually distinguishable from the preceding state, rather than the moment of the preliminary preparatory movement. To evaluate inter-annotator agreement, we utilized a temporal tolerance window. An agreement was considered valid if the transition timestamps provided by the three independent annotators fell within a 0.5-second margin (15 frames), in which case the average timestamp was adopted. In cases where the variance exceeded this threshold or semantic disagreements occurred, a senior researcher acted as an arbitrator to review the video frame-by-frame and determine the final deterministic boundary.

📊 Activity-State Statistics

Posture Distribution

Posture	Train	Test	Total
standing	1,341	223	1,564
sitting	1,257	230	1,487
lying	326	92	418

Atomic Action-Term Distribution

Counts in this table are computed by splitting composite labels after the posture term. Multi-action labels therefore contribute to multiple action terms.

Action term	Count
using a phone	448
drinking	384
reading	370
eating	271
writing	217
using a laptop	216
carrying heavy objects	184
taking off jacket	184
putting on a jacket	180
posture only	175
headache	168
stomachache	162
pouring	149
picking up	129
reaching elevated objects	116
falling	66
washing	58
operating a microwave	52
making coffee	46

🎯 Intended Use

ASTime is intended for research on:

• activity-state transition timestamp localization;
• human activity recognition in long, untrimmed videos;
• robotic perception of human state changes;
• temporal understanding of daily-living, object-manipulation, information-interaction, and health/safety-related activities.

The subject-disjoint split should be used when evaluating generalization to unseen people.

🔒 Privacy and Ethics

We strictly adhered to high ethical standards to protect participant privacy during data collection. The implemented privacy protection framework includes the following core aspects:

• Constructed Simulated Environment: To address privacy concerns, data collection was conducted in a laboratory environment designed to accurately simulate home settings, including bedrooms, rather than in actual private residences.
• Informed Consent and Licensing Transparency: Written informed consent was obtained from all participants, who retained the right to withdraw their data at any time without penalty. Furthermore, participants were fully informed and explicitly agreed to the complete licensing scope and permissible usage boundaries of the ASTime dataset.
• Strict Decency Standards: We enforced rigorous decency protocols, ensuring all participants were fully clothed. The simulated bedroom scenarios were strictly limited to functional activities relevant to elder care monitoring, such as resting or reading in bed, and no intimate or private activities were recorded.
• Comprehensive Data Anonymization: We implemented strict anonymization procedures across all data modalities before analysis. Facial information was blurred to prevent identity identification. All video files underwent complete audio muting to completely remove conversations and ambient sounds. Consequently, all processed data were utilized solely for algorithmic analysis without being publicly linked to participant identities.

(Users should not attempt to identify participants or reverse anonymization. The dataset should be used only in ways consistent with the repository license and data-use terms.)