Dataset Description
Puzzle-STAMPS — A Multimodal Motion–Physiology–Speech Dataset
Known Issues
This page tracks known data quality issues, recording artifacts, and annotation gaps in the Puzzle-STAMPS dataset. Issues are categorized by modality and marked with their current resolution status.
Resolved Fix or workaround is available and documented.
Partial Partially addressed; residual impact documented.
Unresolved Issue is known but not yet corrected.
Documented Inherent limitation; cannot be resolved, impact is documented.
Audio
| # | Issue | Description | Affected sessions / participants | Status |
|---|---|---|---|---|
| A-01 | Cross-talk between lavalier channels | Due to physical proximity of participants and open lavalier microphones, significant acoustic cross-talk is present. This is an inherent property of the recording setup and is noted as a realistic noise condition. Blind source separation methods may be applied by users. | Global | Unresolved |
| A-02 | Interferences | Artifacts in audio recorded from the audio receivers. Audio backups from the audio emitters are available. | Check teams 2, 26, 27 | Partial |
Game-State Logs
| # | Issue | Description | Affected sessions / puzzles | Status |
|---|---|---|---|---|
| G-01 | Timer issues | For some puzzles and sessions the puzzle-box and hint-system had communication issues, creating a delay in displaying relevant hints or skipping puzzles. | Check teams 1, 2, 3, 9, 17, 33 | Unresolved |
| G-02 | Puzzle 8 working randomly | The RFID readers used to detect when Puzzle 8 is solved worked inconsistently across different sessions. | Not quantified at the moment | Documented |
| G-03 | Toolbox logs | The timestamps for when participants were solving the locks to open the toolbox were not recorded | Global | Unresolved |
| G-04 | Puzzle 3 subpuzzles timestamps | Puzzle 3 contains 5 subpuzzles, but timestamps for solving each subpuzzle are not recorded. | Global | Unresolved |
Motion Capture & Localization (OptiTrack / Pozyx RTLS)
| # | Issue | Description | Affected sessions / participants | Status |
|---|---|---|---|---|
| M-01 | Motion tracking cap fell off | Participant's motion tracking cap shifted or fell off during data collection, leading to uncalibrated data. May affect head-orientation-derived measures, particularly forward head orientation estimates. | Check teams 10, 16, 23 | Unresolved |
| M-02 | RTLS tag disconnection | A small number of Pozyx UWB tags became disconnected mid-session, resulting in partial gaps in the room-level localization data for the corresponding participants. The OptiTrack stream remains available as an alternative for head position over the same period. |
Not yet quantified | Partial |
Physiology
| # | Issue | Description | Affected sessions / participants | Status |
|---|---|---|---|---|
| P-01 | ECG artifacts | Quality of the ECG signal is compromised by poor electrode-skin contact. | Check teams 17, 23, 25, 27, 31, 35 | Unresolved |
| P-02 | Logger shutdown | Logger shutdown during data collection, leading to incomplete recordings. | Check teams 17, 23, 25, 27, 31, 35 | Unresolved |
Temporal Alignment
| # | Issue | Description | Affected sessions / participants | Status |
|---|---|---|---|---|
| T-01 | Inter-sensor drift |
The various data streams were recorded on independent systems. Despite the use of a shared NTP
server, a residual misalignment of a few seconds can sometimes be observed between modalities,
caused by network interruptions, variable write latencies, and differences in processing speed.
Impact: For analyses of behaviours that are more global than highly time-precise (e.g., coarse activity, interaction dynamics, session-level statistics), a drift of a few seconds is not expected to affect results. The issue is mainly relevant for fine-grained cross-modal analyses requiring sub-second alignment. Temporary solution: The dataset's built-in cross-modal redundancy (audio captured by both CCTV cameras and lapel microphones, motion recorded independently by the physiological garment IMU, OptiTrack, and Pozyx RTLS) provides anchor points to detect and correct residual misalignment. We provide a correction pipeline based on detected timer events that trigger an audible signal sounds from the timer system as anchors: audio is stretched to match the timer log, video is re-timed to match the aligned audio, and stretching parameters are stored so the alignment can be reproduced, reversed, or applied to other streams (e.g., transcripts). Full details are documented in the dataset README. |
Global (variable magnitude) | Partial |
Video
| # | Issue | Description | Affected sessions / participants | Status |
|---|---|---|---|---|
| V-01 | Variable frame rate & unreliable timestamps | Due to limitations of the Reolink CCTV cameras, raw video recordings have a variable frame rate and unreliable embedded timestamps. We provide both the raw recordings and corrected versions (waiting manual verification) that were automatically re-timed to align with the synchronized lapel audio tracks using cross-correlation-based drift compensation. | Global | Partial |
To report a new issue or provide additional information on an existing one, please contact aat@is.mpg.de.