Dataset Structure Overview

Organization and naming conventions for the UXO Dataset 2024

The UXO Dataset 2024 is organized into a hierarchical structure that separates recordings by target type and includes supporting files for calibration, 3D models, and processing scripts.

Directory Structure

The exported dataset follows this organization:

uxo-dataset2024/
├── recordings/
│   ├── 100lbs/
│   │   ├── <recording-folder>/
│   │   │   ├── aris_raw/
│   │   │   ├── aris_polar/
│   │   │   ├── gopro/
│   │   │   ├── aris_frame_meta.csv
│   │   │   ├── aris_file_meta.yaml
│   │   │   ├── gantry.csv
│   │   │   ├── ar3.csv
│   │   │   └── notes.txt
│   ├── mortar_shell/
│   ├── incendiary/
│   ├── cylinder/
│   └── 100lbs_floor/
├── calibration/
│   └── transforms.yaml
├── 3d_models/
├── scripts/
├── README.md
└── preview.jpg

Naming Conventions

Recording Folders

Recording folders are organized by target type and contain all synchronized data for a single trajectory:

• Target type: The first level of organization (e.g., 100lbs, mortar_shell, incendiary)
• Recording name: Derived from the original ARIS recording filename

Frame Files

All frames within a recording use a consistent 4-digit zero-padded naming scheme:

0000.pgm   # ARIS raw frame at index 0
0000.png   # ARIS polar frame at index 0
0000.jpg   # GoPro frame synchronized to ARIS frame 0
0001.pgm   # ARIS raw frame at index 1
...

The frame index corresponds to the ARIS frame index, which serves as the primary synchronization reference for all other sensors.

Data Organization Principles

ARIS-Centric Synchronization

All data in the dataset is synchronized to the ARIS sonar frames:

• ARIS frames serve as the temporal reference
• GoPro frames are matched to the closest ARIS frame timestamp
• Gantry positions are interpolated to match ARIS frame times
• AR3 poses are calculated from gantry positions and ARIS metadata

Motion Trimming

Recordings are trimmed to include only the active motion period:

1. Motion onset is detected in the ARIS data (using optical flow)
2. GoPro and gantry data are trimmed to match
3. When export_only_with_gopro is enabled, frames without matching GoPro footage are excluded

From release_1_export.py:108:

# If gopro footage is available, only export data when a gopro frame is also available
if trim_from_gopro and gopro_frame is None:
    continue

Target Types

The dataset includes recordings of five distinct UXO target types:

100lbs Bomb

Main target: Standard 100-pound bomb

Mortar Shell

Military mortar shell ordnance

Incendiary

Incendiary device

Cylinder

Cylindrical test object

100lbs Floor

100-pound bomb on floor position

Target types are automatically determined from the notes.txt file in each recording. See the get_target_type() function in release_1_export.py:38.

Export Configuration

The dataset export process is controlled by several configuration parameters in config.yaml:

Parameter	Description	Default
export_gopro_resolution	GoPro resolution (fhd, uhd, etc.)	fhd
export_gopro_format	GoPro image format	jpg
aris_to_polar_image_format	Polar ARIS format	png
export_only_with_gopro	Trim to frames with GoPro data	true

File Count Statistics

A typical recording contains:

• ARIS frames: 100-1000+ frames per recording
• Total dataset: Close to 100 trajectories and over 74,000 frames
• Matched camera frames: Most sonar frames have corresponding GoPro images

When viewing GoPro labels, multiply coordinates by 3 due to the downsampling from UHD to lower resolutions.

Next Steps

File Formats

Learn about the specific file formats used in the dataset

Recording Structure

Understand the contents of each recording folder

Target Details

Explore the UXO targets and their specifications