Skip to main content

​
Overview

BeamFinder outputs all detections to a CSV file at output/detections.csv. Each row represents one detected drone bounding box. The CSV is generated in detect.py:30-48 during the prediction loop.

​
CSV Schema

The output CSV has 7 columns written at detect.py:32: 
image,x_center,y_center,width,height,confidence,class

​
Field Descriptions

​
image
string
Source image filename (basename only, no path).Extracted from Path(r.path).name at detect.py:40.Example: test_001.jpg
​
x_center
float
Bounding box center X coordinate in pixels (absolute, not normalized).Extracted from box.xywh[0][0] and rounded to 2 decimal places at detect.py:44.Range: 0 to image widthExample: 512.34
​
y_center
float
Bounding box center Y coordinate in pixels (absolute, not normalized).Extracted from box.xywh[0][1] and rounded to 2 decimal places at detect.py:44.Range: 0 to image heightExample: 384.67
​
width
float
Bounding box width in pixels.Extracted from box.xywh[0][2] and rounded to 2 decimal places at detect.py:44.Example: 128.45
​
height
float
Bounding box height in pixels.Extracted from box.xywh[0][3] and rounded to 2 decimal places at detect.py:44.Example: 96.23
​
confidence
float
Detection confidence score (0-1).Extracted from box.conf.item() and rounded to 4 decimal places at detect.py:46.Only detections with confidence >= CONF (default 0.4) are included.Range: 0.0 to 1.0Example: 0.8734
​
class
string
Detected object class name.Mapped from class ID using r.names[int(box.cls.item())] at detect.py:47.For BeamFinder, this is always "drone" since there’s only one class.Example: drone

​
Example Output

​
Sample CSV Content

image,x_center,y_center,width,height,confidence,class
test_001.jpg,512.34,384.67,128.45,96.23,0.8734,drone
test_001.jpg,1024.12,720.89,156.78,112.34,0.7621,drone
test_002.jpg,640.00,480.00,140.56,98.12,0.9123,drone
test_003.jpg,800.45,600.23,132.67,105.89,0.6854,drone

​
Multiple Detections Per Image

If an image contains multiple drones, each detection is written as a separate row with the same image value. Example from test_001.jpg above:
  • Detection 1: confidence 0.8734 at (512.34, 384.67)
  • Detection 2: confidence 0.7621 at (1024.12, 720.89)

​
Coordinate System

​
Absolute Pixel Coordinates

All coordinates are in absolute pixels, not normalized [0,1] values.
  • Origin (0, 0) is at the top-left corner of the image
  • X increases rightward
  • Y increases downward

​
Converting to Corner Format

The CSV uses center-width-height format (xywh). To convert to corner coordinates (xyxy): 
x1 = x_center - width / 2   # left
y1 = y_center - height / 2  # top
x2 = x_center + width / 2   # right
y2 = y_center + height / 2  # bottom
Example: 
# From CSV: 512.34, 384.67, 128.45, 96.23
x1 = 512.34 - 128.45/2 = 448.115
y1 = 384.67 - 96.23/2 = 336.555
x2 = 512.34 + 128.45/2 = 576.565
y2 = 384.67 + 96.23/2 = 432.785

​
Converting to Normalized Coordinates

To normalize coordinates for input to other systems: 
x_center_norm = x_center / image_width
y_center_norm = y_center / image_height
width_norm = width / image_width
height_norm = height / image_height

​
Filtering & Thresholds

​
Confidence Threshold

The CONF parameter in detect.py:12 controls which detections are saved: 
CONF = 0.4  # Only save detections with confidence >= 0.4
Adjusting the threshold:
Increase CONF to reduce false positives:
CONF = 0.7  # More conservative
  • Fewer false alarms
  • May miss some true drones
  • Recommended for production deployment

​
Post-Processing Filters

You can apply additional filters after detection: 
import pandas as pd

df = pd.read_csv("output/detections.csv")

# Filter by confidence
df_high_conf = df[df['confidence'] >= 0.8]

# Filter by bounding box size (remove tiny detections)
df_filtered = df[(df['width'] >= 50) & (df['height'] >= 50)]

# Count detections per image
detection_counts = df.groupby('image').size()

​
Output Statistics

The detection script prints a summary when complete (detect.py:50): 
1,247 detections saved to output/detections.csv
This total includes all detections across all processed images that meet the confidence threshold.

​
Annotated Images

In addition to the CSV, BeamFinder saves annotated images with bounding boxes drawn at output/annotated/. These images are generated by the save=True parameter in model.predict() at detect.py:36.

​
Annotation Style

  • Green boxes for detections above confidence threshold
  • Confidence score displayed on each box
  • Class label (“drone”) displayed on each box
Annotated images are for visualization only. All downstream processing should use the CSV file for accurate coordinates.

​
Integration Example

​
Reading Detection Results

import pandas as pd
import matplotlib.pyplot as plt
import matplotlib.patches as patches
from PIL import Image

# Load detections
df = pd.read_csv("output/detections.csv")

# Filter for specific image
image_name = "test_001.jpg"
detections = df[df['image'] == image_name]

# Load and display image
img = Image.open(f"data/images/test/{image_name}")
fig, ax = plt.subplots()
ax.imshow(img)

# Draw bounding boxes
for _, det in detections.iterrows():
    x1 = det['x_center'] - det['width'] / 2
    y1 = det['y_center'] - det['height'] / 2
    
    rect = patches.Rectangle(
        (x1, y1), det['width'], det['height'],
        linewidth=2, edgecolor='green', facecolor='none'
    )
    ax.add_patch(rect)
    
    # Add confidence label
    ax.text(x1, y1 - 5, f"{det['confidence']:.3f}",
            color='green', fontsize=10, weight='bold')

plt.show()

​
Converting for THz Beam Steering

For THz beam steering applications, you may need to convert pixel coordinates to angular positions: 
import numpy as np

def pixel_to_angle(x_center, y_center, image_width, image_height, fov_horizontal, fov_vertical):
    """
    Convert pixel coordinates to azimuth/elevation angles.
    
    Args:
        x_center, y_center: Detection center in pixels
        image_width, image_height: Image dimensions
        fov_horizontal, fov_vertical: Camera field of view in degrees
    
    Returns:
        azimuth, elevation: Angles in degrees from image center
    """
    # Normalize to [-0.5, 0.5]
    x_norm = (x_center / image_width) - 0.5
    y_norm = (y_center / image_height) - 0.5
    
    # Convert to angles
    azimuth = x_norm * fov_horizontal
    elevation = -y_norm * fov_vertical  # Negative because y increases downward
    
    return azimuth, elevation

# Example usage
for _, det in detections.iterrows():
    az, el = pixel_to_angle(
        det['x_center'], det['y_center'],
        1920, 1080,  # Image dimensions
        60, 40       # Camera FOV (degrees)
    )
    print(f"Drone at azimuth={az:.2f}°, elevation={el:.2f}°")

Build docs developers (and LLMs) love

Get started for freeTalk to us