YOLOv8s_EO_Drone_Detection

YOLOv8s – EO Drone Detection (IRIS Benchmark)

Overview

This model is a YOLOv8s object detection model trained as part of the IRIS EO Drone Detection Benchmark.

It represents a baseline model within a controlled evaluation of multiple detection architectures under identical conditions. The purpose of this model is to capture behavior under real-world constraints such as long-range detection, small object scale, and environmental variability, not to represent peak or optimized performance.

---

Benchmark Reference

This model was trained and evaluated within the IRIS EO Drone Detection Benchmark.

👉 Primary Benchmark (Methodology, Evaluation, Comparison):
[Benchmark]

The benchmark defines:

• dataset construction
• annotation methodology
• evaluation protocol
• cross-architecture comparison

This model should be interpreted only within that context.

---

Model Details

• Architecture: YOLOv8s
• Task: Object Detection (Drone)
• Input Modality: EO imagery (RGB)
• Dataset: IRIS benchmark dataset (~1,000 validated annotations)
• Classes: 1 (drone)

---

Uses

Direct Use

This model is intended for:

• benchmark reference
• comparative evaluation
• analysis of model behavior under constrained datasets

Out-of-Scope Use

This model is not intended for:

• production deployment without further validation
• safety-critical detection systems
• environments significantly different from the benchmark dataset

---

Training Details

Training Data

The model was trained on a curated EO drone detection dataset derived from the Anti-UAV dataset: https://anti-uav.github.io/dataset/

The dataset was constructed using similarity-based candidate discovery and human-in-the-loop validation. It contains approximately 1,000 validated annotations and is designed to emphasize:

• long-range detection
• small object scale
• environmental variability

Full dataset construction details are available in the benchmark repository.

---

Training Procedure

The model uses a YOLOv8s architecture and was trained under aligned conditions with other evaluated models.

Training and inference were executed using a standardized YOLO-based interface for reproducibility and consistency.

IRIS is not tied to a specific model framework. This model is one artifact produced within a broader workflow for dataset development, evaluation, and architecture comparison.

---

Evaluation

This model was evaluated as part of the IRIS EO Drone Detection Benchmark.

👉 [Benchmark]

The benchmark includes:

• standardized dataset splits
• aligned preprocessing
• cross-architecture comparison
• both quantitative metrics and qualitative inspection

Evaluation results and analysis are maintained in the benchmark repository as the source of truth.

---

Observed Behavior (Summary)

Within the benchmark:

• Provided consistent and stable detection behavior across conditions
• Served as a reliable baseline for comparison against other architectures
• Demonstrated reduced sensitivity at long range and small object scale relative to transformer-based approaches

---

Limitations

• Dataset size is intentionally constrained (~1,000 annotations)
• Performance reflects early-stage dataset development
• May underperform in:
  • long-range detection scenarios
  • low visibility conditions
  • unseen environments

Model behavior is highly dependent on operating conditions and dataset composition.

---

License

This model is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.

---

About IRIS

IRIS is a visual intelligence platform focused on:

• similarity-driven dataset development
• human-in-the-loop validation
• controlled model comparison across architectures
• lifecycle-aware computer vision systems

This model represents one artifact within that workflow.

Check out the IRIS webpage for all the latest news and updates!

• Hugging Face Models → [HF]
• Case Studies → [Here]