Skip to main content

​
Overview

BeamFinder is a specialized detection pipeline that uses YOLO26s to locate drones in images and outputs bounding box coordinates for THz (terahertz) beam steering applications. Built as part of a research study on line-of-sight beam steering for THz communication, BeamFinder fine-tunes state-of-the-art object detection models on drone imagery and provides precise spatial coordinates for targeting systems.

Quick Start

Get detecting drones in under 5 minutes with our quickstart guide

Installation

Complete installation and setup instructions for training and inference

Training Guide

Fine-tune YOLO26s on custom drone datasets with optimized hyperparameters

Detection Guide

Run inference and export bounding box coordinates to CSV

​
How It Works

BeamFinder operates in two phases:
1

Training Phase

Fine-tunes YOLO26s on the DeepSense Scenario 23 drone dataset (11,387 annotated images) to learn drone-specific features. The pretrained COCO model has no “drone” class, so this fine-tuning step is essential.
2

Detection Phase

Runs the fine-tuned model on test images and outputs normalized bounding box coordinates (x_center, y_center, width, height) to CSV format for downstream beam steering controllers.

​
Key Features

​
YOLO26s Architecture

BeamFinder uses the latest YOLO26s model from Ultralytics, providing state-of-the-art detection accuracy with real-time inference speeds. The model is optimized for single-class detection (drones) with custom training augmentations.

​
Production-Ready Pipeline

  • CSV Export: Outputs detection coordinates in a structured format ready for beam steering systems
  • Batch Inference: Process multiple images efficiently with configurable batch sizes
  • Confidence Filtering: Configurable confidence threshold (default 0.4) to control precision/recall tradeoff
  • Annotated Outputs: Generates visualizations with bounding boxes overlaid on images

​
Optimized for A100 GPUs

The training pipeline includes A100-specific optimizations: 
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
torch.backends.cudnn.benchmark = True
With compile=True for torch.compile support (10-30% faster training on PyTorch 2.x).

​
Dataset

BeamFinder is trained on DeepSense Scenario 23, a comprehensive drone detection dataset:
  • 11,387 total images with YOLO-format bounding box annotations
  • 7,970 training images (70%)
  • 1,708 validation images (15%)
  • 1,709 test images (15%)
  • 960×540 resolution (16:9 aspect ratio)
  • Single class: drone
Images come from 51 different capture sessions with varying lighting conditions, backgrounds, and drone positions. The dataset is shuffled before splitting to ensure diverse representation across all sets.

​
Output Format

Detections are saved to output/detections.csv with the following structure:
ColumnTypeDescription
imagestringSource image filename
x_centerfloatBounding box center X coordinate (pixels)
y_centerfloatBounding box center Y coordinate (pixels)
widthfloatBounding box width (pixels)
heightfloatBounding box height (pixels)
confidencefloatDetection confidence score (0-1)
classstringDetected object class (“drone”)

​
Performance

After 100 epochs of training with optimized hyperparameters:
  • Image size: 960×540 (rectangular training with rect=True)
  • Batch size: Automatic (90% GPU memory utilization)
  • Data augmentation: Rotation (±15°), vertical flip (50%), scale (0.9), translate (0.2)
  • Learning rate: Cosine annealing scheduler
  • Early stopping: 20 epochs patience
The COCO-pretrained YOLO26s model does not include a “drone” class. Fine-tuning on drone-specific data is required for accurate detection.

​
Use Cases

  • THz Beam Steering: Provide real-time target coordinates for terahertz communication systems
  • Counter-UAS Systems: Detect and track unauthorized drones in restricted airspace
  • Research: Study line-of-sight communication challenges with mobile aerial targets
  • Dataset Generation: Create annotated datasets for downstream ML tasks

​
Next Steps

Run Your First Detection

Follow the quickstart to detect drones in under 5 minutes

Set Up Your Environment

Install dependencies and configure your training environment

Build docs developers (and LLMs) love

Get started for freeTalk to us