Skip to main content

Detection/Inference

This guide covers how to use the trained model for detection and inference on spectrogram images.

Repository: GitHub
Pre-trained Models: Google Drive

Detection Overview​

Detection involves:

  1. Loading a trained model
  2. Processing input images
  3. Running inference
  4. Post-processing detections
  5. Saving or visualizing results

Basic Detection​

Simple Detection Command​

python detect.py \
    --weights runs/train/exp28/weights/best.pt \
    --source path/to/images \
    --imgsz 512

This will:

  • Load the trained model
  • Process images from the source directory
  • Save detection results to runs/detect/exp/

Detection Parameters​

Required Parameters​

--weights: Path to model weights file

--weights runs/train/exp28/weights/best.pt

--source: Input source (file, directory, URL, or webcam)

--source path/to/images          # Directory
--source path/to/image.png       # Single image
--source 0                       # Webcam
--source https://example.com/img.jpg  # URL

Image Processing Parameters​

--imgsz: Inference image size (must match training size)

--imgsz 512  # 512×512 pixels

--conf-thres: Confidence threshold (0.0-1.0)

--conf-thres 0.25  # Default: 0.25
--conf-thres 0.5   # Higher confidence
--conf-thres 0.932 # Very high confidence (100% precision)

--iou-thres: Non-Maximum Suppression IoU threshold

--iou-thres 0.45  # Default: 0.45

--max-det: Maximum detections per image

--max-det 1000  # Default: 1000

Output Parameters​

--save-img: Save detection result images

--save-img true

--save-txt: Save detection labels in YOLO format

--save-txt

--save-conf: Include confidence scores in saved labels

--save-txt --save-conf

--project: Project directory for saving results

--project runs/detect

--name: Experiment name

--name exp

Visualization Parameters​

--line-thickness: Bounding box line thickness

--line-thickness 3  # Default: 3

--hide-labels: Hide class labels on images

--hide-labels

--hide-conf: Hide confidence scores on images

--hide-conf

Advanced Parameters​

--classes: Filter detections by class IDs

--classes 0 2 3  # Only detect classes 0, 2, 3

--agnostic-nms: Class-agnostic Non-Maximum Suppression

--agnostic-nms

--augment: Augmented inference (TTA - Test Time Augmentation)

--augment

--half: Use FP16 half-precision inference

--half  # Faster, less accurate

--device: Device selection

--device 0        # GPU 0
--device 0,1,2,3  # Multiple GPUs
--device cpu      # CPU

Detection Examples​

Example 1: Basic Detection​

python detect.py \
    --weights runs/train/exp28/weights/best.pt \
    --source data/test_images \
    --imgsz 512 \
    --conf-thres 0.25 \
    --save-img true

Example 2: High Confidence Detection​

For high-precision detection (100% precision at 0.932 confidence):

python detect.py \
    --weights runs/train/exp28/weights/best.pt \
    --source data/test_images \
    --imgsz 512 \
    --conf-thres 0.932 \
    --save-img true

Example 3: Save Labels​

Save detection results in YOLO format:

python detect.py \
    --weights runs/train/exp28/weights/best.pt \
    --source data/test_images \
    --imgsz 512 \
    --save-txt \
    --save-conf

Example 4: Filter by Class​

Detect only specific signal types:

python detect.py \
    --weights runs/train/exp28/weights/best.pt \
    --source data/test_images \
    --imgsz 512 \
    --classes 9 10  # LFM (9) and FMCW (10)

Example 5: Augmented Inference​

Use Test Time Augmentation for better accuracy:

python detect.py \
    --weights runs/train/exp28/weights/best.pt \
    --source data/test_images \
    --imgsz 512 \
    --augment

Detection Output​

Output Structure​

Detection results are saved to:

runs/detect/exp/
├── image1.png      # Detection result image
├── image2.png
├── ...
└── labels/         # If --save-txt
    ├── image1.txt  # Detection labels
    ├── image2.txt
    └── ...

Detection Format​

Image Output:

  • Original image with bounding boxes
  • Class labels and confidence scores
  • Color-coded by class

Label Output (YOLO format):

class_id x_center y_center width height confidence

Example:

0 0.5 0.5 0.3 0.4 0.95
9 0.2 0.3 0.15 0.2 0.87

Detection Results​

Sample Detection Images​

The model successfully detects various signal types:

Detection Example 1

Multiple signal types detected in a single spectrogram

Detection Example 2

Various signal types with different confidence levels

Detection Example 3

High-confidence detections in complex spectrogram environments

Detection Example 4

Detection in dense signal environments with multiple overlapping signals

Performance Considerations​

Inference Speed​

  • Image Size: Larger images = slower inference
  • Batch Size: Process multiple images for better GPU utilization
  • Device: GPU is much faster than CPU
  • Half Precision: Use --half for faster inference (slight accuracy loss)

Memory Usage​

  • Image Size: Larger images require more GPU memory
  • Batch Processing: Process images one at a time for low memory usage
  • Model Size: Larger models require more memory

Accuracy vs Speed Trade-off​

  • High Confidence (0.9+): Slower (fewer detections), very accurate
  • Medium Confidence (0.25-0.5): Balanced speed and accuracy
  • Low Confidence (<0.25): Faster (more detections), may include false positives

Confidence Threshold Selection​

Based on the Precision-Confidence curve:

Precision-Confidence Curve

  • 0.932: 100% precision (very reliable, fewer detections)
  • 0.6-0.8: High precision (>80%), good balance
  • 0.25-0.5: Moderate precision, more detections
  • <0.25: Lower precision, many detections (may include false positives)

Best Practices​

For High Precision​

Use high confidence threshold:

--conf-thres 0.932  # 100% precision

For High Recall​

Use lower confidence threshold:

--conf-thres 0.25  # More detections

For Balanced Performance​

Use medium confidence threshold:

--conf-thres 0.5  # Good balance

Batch Processing​

Process entire directories:

--source path/to/image/directory

Save Results​

Always save results for analysis:

--save-img true --save-txt --save-conf

Troubleshooting​

No Detections​

  • Lower confidence threshold
  • Check image preprocessing
  • Verify model was trained on similar data
  • Check image size matches training size

Too Many False Positives​

  • Increase confidence threshold
  • Adjust IoU threshold
  • Use class filtering
  • Check model training quality

Slow Inference​

  • Use GPU instead of CPU
  • Reduce image size
  • Use half precision (--half)
  • Process in batches