DARPA PHASE 1+2+3
Multi-modal casualty simulation dataset collected from UGV (ground) platforms across multiple field exercises and phases. The repository bundles casualty video, ground/medic camera footage, still imagery, timestamped manikin vitals, and per-casualty trauma annotations.
Dataset Structure
├── DTC_Trauma_Phase1and2.csv Trauma labels for Phase 1 & 2 (keyed by video_id)
├── DTC_Trauma_Sheet_Phase_3.csv Trauma labels for Phase 3 (keyed by casualty clip path)
├── PHASE_1/
│ └── P1_C{n}_{seg}.mp4 Per-casualty video segments (Casualty 1–3, segments a–l)
├── PHASE_2/
│ └── P2D2_G1_S{n}.MP4 Ground platform video sessions
└── PHASE_3/
└── UGV/ Ground (unmanned ground vehicle) platform
├── Field 2/
│ ├── Field 2 Manikin Timestamped Vitals/ (Casualty 6–11 vitals CSVs)
│ ├── Camera Snapshots/
│ │ ├── UGV1(iphone) Casualty & scene stills (.png)
│ │ ├── UGV2(GoPro) Casualty & scene stills
│ │ └── UGV3(DSLR) Casualty & scene stills
│ ├── UGV2/
│ │ ├── casualty_6 … casualty_11 Per-casualty MP4 clips
│ │ └── Medic View/ Body-worn medic camera, per casualty
│ └── UGV4/
│ ├── Camcorder1_23229 Per-casualty camcorder MP4s
│ └── Camcorder2_23230 Per-casualty camcorder MP4s
└── Field 3/
├── Field 3 Manikin Timestamped Vitals/ (Casualty 12–19 vitals CSVs)
├── Camera Snapshots/
│ └── UGV2(GoPro) Casualty & scene stills
├── UGV2/
│ ├── casualty_12_14 … casualty_19 Per-casualty MP4 clips
│ └── Medic View/ Body-worn medic camera, per casualty
└── UGV4/
├── Camcorder1_23229
└── Camcorder2_23230
Annotations
Two top-level CSVs provide per-casualty trauma labels:
| File | Key column | Label columns |
|---|---|---|
DTC_Trauma_Phase1and2.csv |
video_id (e.g. P1_C1_a) |
trauma_head, trauma_torso, trauma_arm, trauma_leg, severe_hemorrhage, motor_alertness |
DTC_Trauma_Sheet_Phase_3.csv |
casualty_id (relative clip path, e.g. UGV/Field 2/UGV2/casualty_6/P3D2_GA32_S1.mp4) |
trauma_head, trauma_torso_front, trauma_torso_back, trauma_arm_right, trauma_arm_left, trauma_leg_right, trauma_leg_left |
Note: In videos with multiple casualties, the casualty closer to the camera is taken as the query subject.
Label legends
Phase 1 & 2 (DTC_Trauma_Phase1and2.csv) — per-region injury code:
| Value | Meaning |
|---|---|
0 |
Normal / no injury |
1 |
Wound |
2 |
Amputation (arms and legs only) |
severe_hemorrhage and motor_alertness are binary flags (0/1).
Phase 3 (DTC_Trauma_Sheet_Phase_3.csv) — per-region injury code:
| Value | Meaning |
|---|---|
0 |
Amputation |
1 |
Open wound |
2 |
Closed wound |
3 |
Burn |
4 |
Not testable |
Platforms
| Platform | Type | Phases / Fields | Sensors / Devices |
|---|---|---|---|
| UGV | Ground | Phase 2, Phase 3 (Field 2, 3) | iPhone & DSLR snapshots, GoPro, Camcorders, Medic cam |
Naming Convention
- Phase/Day:
P1,P2D2,P3D2,P3D3— Phase, Day - Casualty:
C1–C3(Phase 1) /casualty_6–casualty_19(Phase 3) - Camera positions:
CA,CB,GA,GB,G1,M(Medic) - Segment / Session:
_a–_l(Phase 1 segments),S1,S10(sessions)
Example filenames
P1_C1_a.mp4— Phase 1, Casualty 1, segment aP2D2_G1_S10.MP4— Phase 2 Day 2, ground camera G1, session 10P3D2_GA32_S1.mp4— Phase 3 Day 2, casualty clip (Field 2, Casualty 6)Casualty6_timestamped_vitals_P3D2_CA36_S1.csv— Vitals for Casualty 6, Phase 3 Day 2Field 2_Cas 10_1.png— Snapshot of Casualty 10, Field 2
Modalities
- Video (MP4/MOV) — Phase 1/2 casualty clips, UGV ground teleoperation, and body-worn medic camera
- Images (PNG/JPG) — Stills from iPhone, GoPro, and DSLR capturing casualties and scene context
- Vitals (CSV) — Timestamped manikin physiological data (simulated patient vitals)
- Annotations (CSV) — Per-casualty trauma severity / condition labels
Casualties by Field
| Phase / Field | Casualties | Platforms |
|---|---|---|
| Phase 1 | 1–3 | Casualty video |
| Phase 2 | — | UGV (ground) |
| Phase 3 / Field 2 | 6–11 | UGV |
| Phase 3 / Field 3 | 12–19 | UGV |
Basic Usage (Hugging Face Hub)
Download the full repository snapshot:
from huggingface_hub import snapshot_download
local_dir = snapshot_download(
repo_id="vedkdev/DARPA-PHASE-3",
repo_type="dataset",
)
print(local_dir)
Grab a single file (e.g. the Phase 3 label sheet) without cloning everything:
from huggingface_hub import hf_hub_download
csv_path = hf_hub_download(
repo_id="vedkdev/DARPA-PHASE-3",
repo_type="dataset",
filename="DTC_Trauma_Sheet_Phase_3.csv",
)
Load the trauma annotations and resolve clip paths:
import os
import pandas as pd
labels = pd.read_csv(os.path.join(local_dir, "DTC_Trauma_Sheet_Phase_3.csv"))
# casualty_id is a relative path under PHASE_3/
labels["clip_path"] = labels["casualty_id"].apply(
lambda p: os.path.join(local_dir, "PHASE_3", p)
)
print(labels[["casualty_id", "trauma_head", "clip_path"]].head())
Phase 1 & 2 labels join on video_id, which matches the MP4 stem under PHASE_1/:
p12 = pd.read_csv(os.path.join(local_dir, "DTC_Trauma_Phase1and2.csv"))
p12["clip_path"] = p12["video_id"].apply(
lambda vid: os.path.join(local_dir, "PHASE_1", f"{vid}.mp4")
)
The media files are tracked with Git LFS — run git lfs install first if you prefer a full git clone over snapshot_download.
Intended Uses
- Multi-view casualty detection and tracking
- Triage scene understanding from ground perspectives
- Multi-modal fusion (video + vitals + still imagery)
- Trauma severity classification and medic action recognition
- Simulated trauma response research
License
Restricted — contact dataset maintainers for usage terms.
- Downloads last month
- -