Dataset Card for X-Ego-CS

Links:

• Paper
• Github Codebase
• Homepage (comming soon)

Cross-Ego Demo (Pistol Round)

Note: This demo concats videos in a grid. The original datasets videos are from individual player POV recording.

Dataset Summary

X-Ego-CS is a multi-agent gameplay video dataset for cross-egocentric multi-agent video understanding in Counter-Strike:2. It contains 124 hours of synchronized first-person gameplay footage captured from 45 professional-level Counter-Strike 2 matches. Each match includes multi-player egocentric video streams (POVs from all players) and corresponding state-action trajectories, enabling the study of team-level tactical reasoning and situational awareness from individual perspectives.

The dataset was introduced in the paper:

X-Ego: Acquiring Team-Level Tactical Situational Awareness via Cross-Egocentric Contrastive Video Representation Learning
Yunzhe Wang, Soham Hans, Volkan Ustun
University of Southern California, Institute for Creative Technologies (2025)
arXiv:2510.19150

X-Ego-CS supports research on multi-agent representation learning, egocentric video modeling, team tactic analysis, and AI-human collaboration in complex 3D environments.

---

Dataset Structure

Data Fields

Segment Info

• idx — Row index (unique for each segment)
• partition — Dataset split label (e.g., train/test/val)
• seg_duration_sec — Duration of the segment in seconds
• start_tick, end_tick, prediction_tick — Game tick indices for start, end, and prediction points
• start_seconds, end_seconds, prediction_seconds — Corresponding timestamps in seconds
• normalized_start_seconds, normalized_end_seconds, normalized_prediction_seconds — Time values normalized to a [0–1] scale for model input

Match Metadata

• match_id — Unique identifier of the match
• round_num — Match round number
• map_name — Name of the game map (e.g., de_mirage)

Player States (for player_0 → player_9)

• player_{i}_id — Unique identifier (e.g., Steam ID)
• player_{i}_name — In-game player name
• player_{i}_side — Team side (t for Terrorist, ct for Counter-Terrorist)
• player_{i}_X, player_{i}_Y, player_{i}_Z — Player’s position coordinates (normalized or map-based)
• player_{i}_place — Named location or area on the map (e.g., CTSpawn, SideAlley)

File Structure

data/
├── demos/                       # Raw .dem files (by match)
│   └── <match_id>.dem
├── labels/                      # Global label datasets
│   ├── enemy_location_nowcast_s1s_l5s.csv
│   └── teammate_location_nowcast_s1s_l5s.csv
├── metadata/                    # Match / round metadata
│   ├── matches/
│   │   └── <match_id>.json
│   └── rounds/
│       └── <match_id>/
│           └── round_<nn>.json
├── trajectories/                # Player movement trajectories
│   └── <match_id>/
│       └── <player_id>/
│           ├── round_<nn>.csv
│           └── ...
└── videos/                      # Player POV recordings
    └── <match_id>/
        └── <player_id>/
            ├── round_<nn>.mp4
            └── ...

---

Dataset Creation

Curation Rationale

The dataset was designed to study cross-perspective alignment in team-based tactical games where each player’s view provides only a partial observation of the environment. Synchronizing multiple first-person streams allows for modeling shared situational awareness and implicit coordination—key ingredients in human team intelligence.

Source Data

• Game: Counter-Strike 2 (Valve Corporation) in-game demo replay recording. Downloaded from top elo-leaderboard from Faceit.com
• Recording setup: Screen capture of first-person gameplay, synchronized across all agents using timestamp alignment
• Annotations: Automatically generated state-action trajectories derived from server replay data

---

Dataset Statistics

• Total hours: 124
• Total matches: 45
• Agents per match: 10 (5 per team)
• Frame rate: 30 fps
• Video resolution: 1920×1080

---

Citation

If you use this dataset, please cite the following paper:

@article{wang2025x,
  title={X-Ego: Acquiring Team-Level Tactical Situational Awareness via Cross-Egocentric Contrastive Video Representation Learning},
  author={Wang, Yunzhe and Hans, Soham and Ustun, Volkan},
  journal={arXiv preprint arXiv:2510.19150},
  year={2025}
}