Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/dfki-ric/uxo-dataset2024/llms.txt

Use this file to discover all available pages before exploring further.

The UXO Dataset 2024 includes recordings of five distinct unexploded ordnance (UXO) target types, each with known 3D position, orientation, and available 3D models.

​
Target Types

The dataset includes five UXO targets recorded in a controlled experimental environment:

100lbs

Standard 100-pound aircraft bombPosition: (1.251, 1.593, -1.490) mPrimary target with the most recordings

Mortar Shell

Military mortar shell ordnancePosition: (1.264, 1.437, -1.519) mSmaller cylindrical munition

Incendiary

Incendiary devicePosition: (1.235, 1.409, -1.531) mSpecialized munition type

Cylinder

Cylindrical test objectPosition: (1.278, 1.400, -1.461) mReference geometry for comparison

100lbs Floor

100-pound bomb in floor positionPosition: (1.174, 1.228, -2.241) mSame bomb at a different location

​
Transform Hierarchy

All target positions and sensor poses are defined in a calibrated transform tree stored in calibration/transforms.yaml.

​
Coordinate Frames

The dataset uses the following coordinate frame hierarchy: 
world (tank fixed frame)
├── setup/portal_crane (gantry crane base)
│   └── setup/ar3 (robot arm base)
│       ├── setup/aris (ARIS mounting point)
│       │   └── setup/sonar (ARIS acoustic center)
│       └── setup/gopro_mounting_hole
│           └── setup/gopro_mounting_axis
│               └── setup/camera (GoPro optical center)
└── target/* (UXO target frames)

​
Setup Transforms

These transforms define the sensor mounting geometry:
Frame: setup/ar3Parent: setup/portal_craneTransform:
translation:
  x: 0.0
  y: 0.174
  z: -0.338
rotation:  # Quaternion (x, y, z, w)
  x: 0.0
  y: 1.0
  z: 0.0
  w: 0.0
This transform positions the AR3 robot arm base relative to the gantry crane.
Frame: setup/arisParent: setup/ar3Transform:
translation:
  x: 0.0
  y: 0.0
  z: 0.0
rotation:
  x: 0.0
  y: 0.0
  z: 0.0
  w: 1.0
The ARIS is mounted directly on the AR3 (identity transform).
Frame: setup/sonarParent: setup/arisTransform:
translation:
  x: -0.108725
  y: 0.0
  z: 0.0
rotation:
  x: 0.5
  y: 0.5
  z: 0.5
  w: 0.5
Offset from the ARIS mounting point to the acoustic center, including 90° rotations to align coordinate frames.
Frame: setup/cameraParent: setup/ar3 (through intermediate mounting frames)Transform chain:
# AR3 to mounting hole
setup/gopro_mounting_hole:
  translation:
    x: 0.1265
    y: 0.023
    z: -0.1055
  rotation:
    x: 0.0
    y: 0.0
    z: 0.0
    w: 1.0

# Mounting hole to mounting axis
setup/gopro_mounting_axis:
  translation:
    x: 0.02125
    y: 0.0
    z: -0.0075
  rotation:
    x: 0.0
    y: 0.0
    z: 0.0
    w: 1.0

# Mounting axis to camera optical center
setup/camera:
  translation:
    x: 0.01925
    y: -0.0195
    z: -0.0415
  rotation:
    x: 0.5
    y: 0.5
    z: 0.5
    w: 0.5
The GoPro mounting includes multiple frames to account for the physical mounting bracket geometry.
Full setup transforms in calibration/transforms.yaml:1-71.

​
Target Positions

All target positions are defined relative to the world (tank) coordinate frame.

​
100lbs Bomb

Frame ID: target/100lbs Position: 
translation:
  x: 1.251079
  y: 1.592704
  z: -1.489828
rotation:
  x: 0.0
  y: 0.0
  z: 0.0
  w: 1.0
Recording folder: recordings/100lbs/ Description: Standard 100-pound aircraft bomb, the primary target in the dataset.

​
Mortar Shell

Frame ID: target/mortar_shell Position: 
translation:
  x: 1.263516
  y: 1.4371
  z: -1.51891
rotation:
  x: 0.0
  y: 0.0
  z: 0.0
  w: 1.0
Recording folder: recordings/mortar_shell/ Description: Military mortar shell with cylindrical geometry.

​
Incendiary

Frame ID: target/incendiary Position: 
translation:
  x: 1.2351
  y: 1.409233
  z: -1.530784
rotation:
  x: 0.0
  y: 0.0
  z: 0.0
  w: 1.0
Recording folder: recordings/incendiary/ Description: Incendiary device munition.

​
Cylinder

Frame ID: target/cylinder Position: 
translation:
  x: 1.277832
  y: 1.399541
  z: -1.461439
rotation:
  x: 0.0
  y: 0.0
  z: 0.0
  w: 1.0
Recording folder: recordings/cylinder/ Description: Simple cylindrical reference object for comparison with UXO targets.

​
100lbs Floor

Frame ID: target/100lbs_floor Position: 
translation:
  x: 1.174003
  y: 1.227536
  z: -2.2411
rotation:
  x: 0.0
  y: 0.0
  z: 0.0
  w: 1.0
Recording folder: recordings/100lbs_floor/ Description: The same 100lbs bomb positioned on the tank floor, at a different depth and location. Full target transforms in calibration/transforms.yaml:72-127.

​
3D Models

Textured 3D models are available for all targets in the 3d_models/ directory.

​
Model Formats

The models are provided in standard formats:
  • Mesh files: .obj, .ply, or similar formats
  • Texture files: .jpg, .png for texture maps
  • Metadata: Scale and alignment information

​
Model Generation

The 3D models were created using photogrammetry (Agisoft Metashape) from optical imagery of the targets.
Metashape project files are excluded from the export. Only the final textured models are included.
From release_1_export.py:175-181: 
# Copy 3d models
print('Copying 3d models...')
model_dir = os.path.join(export_dir, '3d_models/')
shutil.copytree(os.path.join(data_root, '../3d_models'), 
                model_dir, 
                dirs_exist_ok=True, 
                ignore=lambda src, names: [x for x in names if 'metashape' in x])

​
Model Coordinate Frames

The 3D models are aligned to their respective target coordinate frames defined in transforms.yaml. This allows:
  • Direct overlay of models on sensor data
  • Accurate range and bearing calculations
  • Ground truth for target detection and localization

​
Working with Transforms

The dataset includes a transform manager for working with the calibration data.

​
Python Example

import numpy as np
from scipy.spatial.transform import Rotation as R
from dataset.calibration.tf_demo.transforms import get_tf_manager

# Load the transform manager
tm = get_tf_manager()

# Get transform from sonar to a target
T_sonar_to_100lbs = tm.get_transform("target/100lbs", "setup/sonar")

# Extract position and rotation
target_pos = T_sonar_to_100lbs[:3, 3]
target_rot = R.from_matrix(T_sonar_to_100lbs[:3, :3])

print(f"Target position in sonar frame: {target_pos}")
print(f"Target orientation: {target_rot.as_euler('xyz', degrees=True)}°")

# Calculate range and bearing
range_to_target = np.linalg.norm(target_pos)
bearing = np.arctan2(target_pos[1], target_pos[0]) * 180 / np.pi

print(f"Range: {range_to_target:.2f} m")
print(f"Bearing: {bearing:.2f}°")

​
Dynamic Transforms

The gantry position changes with each frame, creating dynamic transforms: 
import pandas as pd

# Load gantry data
gantry = pd.read_csv('recording_001/gantry.csv', index_col='aris_frame_idx')

# Set gantry position for a specific frame
frame_idx = 42
gantry_pos = gantry.loc[frame_idx][['x', 'y', 'z']].values

# Create transform
T_world_to_gantry = np.eye(4)
T_world_to_gantry[:3, 3] = gantry_pos

# Add to transform manager
tm.add_transform("setup/portal_crane", "world", A2B=T_world_to_gantry)

# Now you can get the sonar position in world frame
T_world_to_sonar = tm.get_transform("setup/sonar", "world")
sonar_pos = T_world_to_sonar[:3, 3]
This is exactly how the AR3 poses are calculated in release_1_export.py:18-36.

​
Sonar-to-Target Geometry

With the calibrated transforms and known target positions, you can calculate the exact geometry for each frame:

​
Range and Bearing Calculation

import numpy as np
import pandas as pd
from dataset.calibration.tf_demo.transforms import get_tf_manager

def calculate_target_geometry(recording_dir: str, frame_idx: int, target_name: str):
    """Calculate range and bearing to target for a specific frame."""
    
    # Load data
    gantry = pd.read_csv(f"{recording_dir}/gantry.csv", index_col='aris_frame_idx')
    frame_meta = pd.read_csv(f"{recording_dir}/aris_frame_meta.csv", 
                            index_col='FrameIndex')
    
    # Get gantry position
    gantry_pos = gantry.loc[frame_idx][['x', 'y', 'z']].values
    
    # Get ARIS orientation
    aris_angles = frame_meta.loc[frame_idx][['SonarPan', 'SonarTilt', 'SonarRoll']].values
    
    # Setup transform manager
    tm = get_tf_manager()
    
    # Add gantry transform
    T_gantry = np.eye(4)
    T_gantry[:3, 3] = gantry_pos
    tm.add_transform("setup/portal_crane", "world", A2B=T_gantry)
    
    # Get sonar to target transform
    T_sonar_to_target = tm.get_transform(f"target/{target_name}", "setup/sonar")
    
    # Extract position
    target_pos = T_sonar_to_target[:3, 3]
    
    # Calculate range and bearing
    range_m = np.linalg.norm(target_pos)
    bearing_deg = np.arctan2(target_pos[1], target_pos[0]) * 180 / np.pi
    elevation_deg = np.arctan2(target_pos[2], 
                              np.sqrt(target_pos[0]**2 + target_pos[1]**2)) * 180 / np.pi
    
    return {
        'range': range_m,
        'bearing': bearing_deg,
        'elevation': elevation_deg,
        'position': target_pos
    }

# Usage
geometry = calculate_target_geometry('recordings/100lbs/recording_001', 42, '100lbs')
print(f"Range: {geometry['range']:.2f} m")
print(f"Bearing: {geometry['bearing']:.2f}°")
print(f"Elevation: {geometry['elevation']:.2f}°")

​
Ground Truth Labels

The known target positions provide ground truth for:
  • Target detection: True positive validation
  • Localization accuracy: Range/bearing error metrics
  • Tracking performance: Multi-frame association
  • Segmentation: Pixel-wise target masks
  • 3D reconstruction: Point cloud accuracy

​
Target Specifications

​
Physical Dimensions

While exact dimensions depend on the specific munition types, typical characteristics:
  • Type: Aircraft bomb
  • Weight: 100 pounds (~45 kg)
  • Shape: Cylindrical body with tail fins
  • Length: ~0.8-1.0 m (estimated)
  • Diameter: ~0.15-0.20 m (estimated)
Exact dimensions can be measured from the 3D models in the 3d_models/ directory.

​
Experimental Setup

The targets were positioned in a controlled experimental tank:
  • Fixed positions: Targets remain stationary during recording
  • Known geometry: Accurate surveying and calibration
  • Controlled environment: Indoor tank with controlled water conditions
  • Repeatable trajectories: Gantry crane enables precise motion control
Recording setup From the README:
Using an ARIS 3000 imaging sonar, a GoPro Hero 8 and a custom design gantry crane, we recorded close to 100 trajectories and over 74,000 frames of 3 distinct types of UXO in a controlled environment.

​
Next Steps

Dataset Overview

Return to the dataset structure overview

File Formats

Learn about the specific file formats

Recordings

Explore recording folder organization

Build docs developers (and LLMs) love

Get started for freeTalk to us