Raw-Drone Subset Benchmark

• Raw-Drone Subset is a minimally processed segment of multimodal drone data that preserves original sensor recordings without normalization or denoising.
• It supports reproducible research by providing standardized acoustic, imagery, multispectral, and RF data with detailed metadata for benchmarking.
• The subset enables rigorous evaluation in detection, classification, and signal processing pipelines, fostering reliable model development and discovery.

A Raw-Drone Subset is a rigorously defined, minimally processed segment of a multimodal drone dataset that preserves primary sensor recordings prior to feature extraction, normalization, or augmentation. Raw-Drone Subsets serve as canonical ground truth for benchmarking, model development, and data-driven discovery across drone acoustics, imagery, and radio-frequency (RF) signal domains. The subset’s defining principle is fidelity to the original data acquisition—minimal trimming, no denoising, and explicit linkage to fully documented metadata and labeling schemas.

1. Dataset Characterization

Raw-Drone Subsets are crucial for reproducible research and standardized evaluation. They appear in multiple modalities:

• Acoustic: The "Raw-Drone Subset" in "A Multiclass Acoustic Dataset and Interactive Tool for Analyzing Drone Signatures in Real-World Environments" consists of 3,200 16-bit PCM WAV audio clips, totaling 16,000 seconds, each of exactly 5 seconds, sampled at 44.1 kHz (mono). This covers 32 drone categories, each contributing 100 uniformly distributed clips. Recording environments include outdoor rural, urban rooftop, and indoor labs, using a MacBook Air internal microphone. No normalization, filtering, or denoising is applied; these files are manually trimmed to 5 seconds and stored under a standardized directory and naming scheme, with metadata available via JSON manifest, supporting robust downstream tasks and exploration (Wang et al., 5 Sep 2025).
• Imagery: In data-scarce detection, the raw portion can consist of ≈4,000 simulated RGB images (e.g., rendered from 14 CAD UAV models, 64×64 pixels, normalized per channel). Image-based raw subsets are typically batch-packed for generative modeling, with expert-defined tags retained (Li et al., 2021).
• Multispectral/Multisensor: The raw subset in multi-sensor datasets includes synchronized IR (e.g., 320×256 px, 16 bpp) and visible videos (1280×720 px, 50 FPS), as well as directional microphone WAV audio (44.1 kHz, 16 bit), each categorized by distance bins (Close, Medium, Distant) according to DRI/Johnson criteria, and labeled per class and acquisition distance (Svanström et al., 2021).
• RF Domain: Recent RF benchmarks define raw subsets as multi-GB or TB archives of unprocessed I/Q streams (complex float32), with per-device classes, multi-band captures (900 MHz, 2.4 GHz, 5.8 GHz), and exacting acquisition metadata. Signal quality is not modified unless specifically post-processed for SNR control; each segment is strictly associated with detailed session metadata, including device, center frequency, gain, and channel conditions (Shi et al., 12 Mar 2025, Rostami et al., 6 Jan 2026).

2. File Specifications and Organization

Data structure in Raw-Drone Subsets is governed by reproducibility, traceability, and interoperability mandates:

Modalities	Format/Specs	Storage/Naming Conventions
Acoustic (audio)	16-bit PCM WAV, 44.1 kHz, mono	audio/raw/<brand_model>/<brand_model>_<idx>.wav
Imagery	PNG/JPG, 64×64 (RGB) or custom sizes	<model>/<img_idx>.png
Video	H.264 MP4 (IR: 320×256, VIS: 1280×720)	<Sensor>_<Class>_<idx>.mp4
RF Signals	Binary float32 I/Q arrays	<MODEL>_<SESSION>.bin, with companion XML/CSV metadata

Metadata mapping is standardized—e.g., JSON manifest for acoustic data, MATLAB .mat GroundTruth objects for video/IR, and CSV/XML for RF signal sessions. Category labels are typically integer-encoded by sorted manufacturer/model.

3. Acquisition Protocols and Metadata

Recording conditions are meticulously logged to enable controlled experiments:

• Acoustic: Drones are hovered at 1–5 meters from the sensor (laptop or directional microphone) in indoor and outdoor environments. Ambient noise is preserved (wind, birds, traffic); the mean SNR is not pre-assessed (Wang et al., 5 Sep 2025).
• Imagery: Simulated images are generated via CAD model rotation, no post-shoot preprocessing is applied, and descriptive tags (model, color, configuration) are retained (Li et al., 2021).
• Multisensor: Cameras and microphones are synchronized, and each clip is labeled with exact class, sensor, and DRI-based distance bin. Environmental and acquisition metadata are included in associated Excel/CSV sheets (Svanström et al., 2021).
• RF: SDR-based receivers (e.g., USRP X310 or B200-mini) capture over variable durations (tens to hundreds of seconds), with antenna, gain, frequency band, and environmental labels embedded in filenames or sidecar XMLs. RF cage and field scenarios yield both low- and high-interference exemplars (Shi et al., 12 Mar 2025, Rostami et al., 6 Jan 2026).

4. Preprocessing, Benchmarks, and Downstream Processing

The unprocessed nature of Raw-Drone Subsets supports rigorous benchmarking and feature extraction. Downstream pipelines are precisely specified:

• Acoustic Feature Extraction: MFCC pipeline using Hanning window
  • $w[n] = \frac{1}{2}[1 - \cos(\frac{2\pi n}{N})]$
  • STFT: $$\mathrm{STFT}_x(t,\omega)=\int_{-\infty}^\infty x(\tau)\,w(\tau-t)\,e^{-j\omega\tau}d\tau$$
  • Mel-scale: $m(f)=2595\log_{10}\left(1 + \frac{f}{700}\right)$
  • Triangular filter bank and DCT for cepstral coefficients (Wang et al., 5 Sep 2025).
• Imagery: Used as-is for training/WGANs. Feature space holes are identified via Topological Data Analysis on last-layer activations (Li et al., 2021).
• RF: Raw I/Q streams are post-processed through STFT (e.g., windowed Hamming filter, $N=256$ to 1024 FFT points), optional bandpass filtering, and dB-scaling. SNR sweeps use additive complex Gaussian noise:
  • $$x_\text{{noisy}}[n] = x[n] + \alpha w[n],\ \alpha = \sqrt{\frac{\sum_n |x[n]|^2}{\sum_n |w[n]|^2} \cdot 10^{-\text{SNR}/10}}$$
  • Fingerprinting slices I/Q into $K$ blocks, extracts $f_k$ via log-magnitude spectrum or cyclostationary statistics (Shi et al., 12 Mar 2025, Rostami et al., 6 Jan 2026).
• Augmentation: Raw-Drone Subset defines ground truth; augmentation is handled in dedicated modules, e.g., for SNR, frequency shift, or interferer injection, with deterministic label adjustment (e.g., bounding-box translation in spectrograms) (Rostami et al., 6 Jan 2026).

5. Benchmark Protocols, Labeling, and Evaluation

Benchmarking leverages the raw subset for classification, detection, and open-set recognition:

Benchmark	Modality	Label Structure	Typical Metric
YOLO Detection	RF/Audiospectro.	YOLO: [class, x_c, y_c, w, h]	Precision, Recall, mAP
Classification	Audio/RF/Image	Integer per-class label	Accuracy, hierarchical score
Open-set Recog.	RF/Image	Known/unknown class splits	Overall, unknown det. acc.

For example, CageDroneRF divides raw and augmented data by train/validation/test folders, stratified across classes and SNR sweeps, with object detection and single-label classification both supported. Metadata is always maintained: timestamp, device ID, acquisition parameters, and augmentation flags (Rostami et al., 6 Jan 2026). In multisensor and acoustic domains, evaluation data splits are typically unconstrained—stratification is recommended by class and acquisition condition (Svanström et al., 2021, Wang et al., 5 Sep 2025).

6. Applications and Significance

Raw-Drone Subsets underpin several distinct methodological and applied research directions:

• Detection Algorithm Benchmarking: Standardized raw audio and RF subsets facilitate robust comparison of detection architectures (e.g., CNNs, YOLO variants, ResNet, GANs) across an unprocessed baseline (Wang et al., 5 Sep 2025, Shi et al., 12 Mar 2025, Rostami et al., 6 Jan 2026).
• Feature-Space Analysis: Raw-image subsets enable deep generative modeling (DCGAN, WGAN), quantifying data space holes and guiding targeted data acquisition, which demonstrably improves deep model performance beyond generic augmentation (Li et al., 2021).
• Signal Robustness: SNR-parameterizable RF raw data enables systematic robustness analysis, including evaluation of performance across SNR sweeps, interferer scenarios, and frequency offsets—crucial for real-world deployments (Shi et al., 12 Mar 2025, Rostami et al., 6 Jan 2026).
• Standardization: Open-source code, public repositories, and canonical directory/labeling schemes enforce interoperability across research groups, supporting reproducible science and meta-analyses (Svanström et al., 2021, Rostami et al., 6 Jan 2026).

7. Access, Licensing, and Tooling

Raw-Drone Subsets are generally released in open-access repositories, with permissive licensing (e.g., CC BY 4.0 or equivalent). Data retrieval, metadata access, and benchmark suite invocation are documented with example commands (e.g., python module calls for SNR estimation, augmentation, or detection/classification benchmarking). Public web portals, interactive visualization tools, and automated API endpoints further facilitate dataset exploration and community benchmarking (Wang et al., 5 Sep 2025, Svanström et al., 2021, Shi et al., 12 Mar 2025, Rostami et al., 6 Jan 2026).