Test Image

Overview

The --test-image flag allows you to test person detection on a static image file before processing live RTSP streams. This is essential for:

Validating detection setup – Verify your model files are loaded correctly
Tuning parameters – Find optimal confidence and area thresholds
Debugging issues – Understand why detections may be failing
Testing hardware – Confirm GPU acceleration is working

Test image mode runs a single detection pass and exits. It does not connect to any RTSP streams.

Basic Usage

uv run main.py --test-image photo.jpg --save image

Output:

Config loaded from: config.cfg
  model_dir   = model
  output_dir  = output
Loading person detection model...
Model loaded: YOLOv4
Confidence threshold: 0.5
Person area threshold: 1000 pixels
Testing with image: photo.jpg
Persons detected: 2
Bounding boxes: [(145, 89, 234, 456, 0.87), (456, 102, 189, 423, 0.72)]
Annotated result saved to: test_result_1741528222.jpg

Step-by-Step Guide

Prepare a test image

Use any JPEG or PNG image containing people. For best results:

Resolution: Similar to your RTSP stream resolution
Lighting: Similar conditions to your deployment environment
Distance: People at similar distances as your camera setup

Example test images:

# Download a sample image
wget https://example.com/sample-crowd.jpg -O test.jpg

# Or use a frame from your stream
ffmpeg -i rtsp://camera.local/stream -frames:v 1 test.jpg

Run detection with default settings

Test with default configuration:

uv run main.py --test-image test.jpg --save image

This uses default thresholds from config.cfg:

confidence_threshold = 0.5
person_area_threshold = 1000

Review the annotated output

Open the generated test_result_*.jpg file:

Green bounding boxes around detected persons
Confidence scores displayed above each box (e.g., “Person 1: 0.87”)
Box dimensions correspond to bounding box size in pixels

Check the console output:

Persons detected: 2
Bounding boxes: [(145, 89, 234, 456, 0.87), (456, 102, 189, 423, 0.72)]

Format: (x, y, width, height, confidence)

Adjust thresholds if needed

If results aren’t as expected, tune parameters:

# Lower confidence to detect more persons
uv run main.py --test-image test.jpg --save image --confidence 0.3

# Higher confidence for fewer false positives
uv run main.py --test-image test.jpg --save image --confidence 0.7

# Filter small detections
uv run main.py --test-image test.jpg --save image --area-threshold 5000

Test with your production settings

Once you find optimal parameters, test them:

uv run main.py --test-image test.jpg --save image \
  --confidence 0.6 \
  --area-threshold 2500 \
  --config production.cfg

If results look good, use the same settings for live streams.

Understanding the Output

Console Output Breakdown

Persons detected: 2
Bounding boxes: [(145, 89, 234, 456, 0.87), (456, 102, 189, 423, 0.72)]
Annotated result saved to: test_result_1741528222.jpg

Bounding box format: (x, y, width, height, confidence)

Field	Description	Example
`x`	Left edge position (pixels)	145
`y`	Top edge position (pixels)	89
`width`	Box width (pixels)	234
`height`	Box height (pixels)	456
`confidence`	Detection confidence (0.0-1.0)	0.87 (87%)

Box area: width × height = 234 × 456 = 106,704 pixels

Annotated Image

The output image test_result_*.jpg shows:

Green rectangles – Bounding boxes around detected persons
Labels – “Person 1: 0.87”, “Person 2: 0.72”, etc.
Original image – Background preserved, detections overlaid

The annotated image is saved with a Unix timestamp in the filename (e.g., test_result_1741528222.jpg) to prevent overwriting previous test runs.

Tuning Detection Parameters

Confidence Threshold Examples

Low (0.3) - High Recall
Default (0.5) - Balanced
High (0.7) - High Precision

uv run main.py --test-image test.jpg --save image --confidence 0.3

Result:

Persons detected: 5
Bounding boxes: [(145, 89, 234, 456, 0.87), (456, 102, 189, 423, 0.72), 
                 (234, 156, 123, 289, 0.45), (678, 234, 98, 234, 0.38), 
                 (890, 123, 156, 345, 0.32)]

Interpretation: More detections, including lower-confidence ones. May include false positives (non-person objects).

uv run main.py --test-image test.jpg --save image

Result:

Persons detected: 3
Bounding boxes: [(145, 89, 234, 456, 0.87), (456, 102, 189, 423, 0.72), 
                 (234, 156, 123, 289, 0.54)]

Interpretation: Balanced accuracy. Good starting point for most use cases.

uv run main.py --test-image test.jpg --save image --confidence 0.7

Result:

Persons detected: 2
Bounding boxes: [(145, 89, 234, 456, 0.87), (456, 102, 189, 423, 0.72)]

Interpretation: Only high-confidence detections. Fewer false positives, but may miss some persons.

Area Threshold Examples

Small (500) - Detect Distant Persons
Medium (2500) - Nearby Persons
Large (10000) - Close-Up Only

uv run main.py --test-image test.jpg --save image --area-threshold 500

Captures small bounding boxes (distant persons, children, partial views).

uv run main.py --test-image test.jpg --save image --area-threshold 2500

Filters out small detections. Good for close-range cameras.

uv run main.py --test-image test.jpg --save image --area-threshold 10000

Only captures large, close-up persons. Ignores background and distant people.

Troubleshooting Detection Issues

No persons detected (0 detections)

Possible causes:

Thresholds too high – Try lowering them:

uv run main.py --test-image test.jpg --save image \
  --confidence 0.2 \
  --area-threshold 100

Model files missing – Check for HOG fallback warning:
```
Warning: YOLO weights not found. Using OpenCV's built-in HOG person detector as fallback.
```
HOG is less accurate than YOLO. Download proper model files.
Image issues – Verify image loaded correctly:
```
file test.jpg  # Should show: JPEG image data
```
Poor image quality – Try with a clearer image or better lighting.

Too many false positives

Solution: Increase thresholds:

uv run main.py --test-image test.jpg --save image \
  --confidence 0.7 \
  --area-threshold 5000

Also check the annotated image – false positives may be objects that resemble humans (mannequins, posters, etc.).

Missing obvious persons

Possible causes:

Confidence too low – Person detected but below threshold
Area too small – Person detected but filtered by area threshold
Model limitations – YOLO/HOG struggle with certain poses or occlusions

Solution: Lower both thresholds:

uv run main.py --test-image test.jpg --save image \
  --confidence 0.3 \
  --area-threshold 500

Bounding boxes in wrong locations

This shouldn’t happen with test images. If you see incorrect boxes:

Check image format – Use standard JPEG/PNG
Verify model files – Re-download YOLO weights
Check OpenCV version – Ensure opencv-contrib-python is installed

python -c "import cv2; print(cv2.__version__)"

Error: Image file not found

Output:

Error: Image file test.jpg not found

Solution: Provide the full or correct relative path:

uv run main.py --test-image /full/path/to/test.jpg --save image
uv run main.py --test-image ./images/test.jpg --save image

Error: Could not load image

Output:

Error: Could not load image test.jpg

Possible causes:

Corrupted image file
Unsupported format
File permissions

Solution:

# Check file integrity
file test.jpg

# Convert to standard JPEG
convert test.png test.jpg

# Fix permissions
chmod 644 test.jpg

Comparing Detection Methods

YOLOv4 vs YOLOv3 vs HOG

YOLOv4 (Recommended)
YOLOv3 (Fallback)
HOG (Fallback)

Accuracy: ★★★★★
Speed: ★★★★☆
GPU Support: Yes

# Ensure model files exist:
ls model/yolov4.weights model/yolov4.cfg model/coco.names

Expected output:

Model loaded: YOLOv4
CUDA available, using GPU for inference

Best for production deployments.

Accuracy: ★★★★☆
Speed: ★★★★☆
GPU Support: YesUsed automatically if YOLOv4 files aren’t found:

# Download YOLOv3 files:
wget https://pjreddie.com/media/files/yolov3.weights -O model/yolov3.weights
wget https://github.com/pjreddie/darknet/raw/master/cfg/yolov3.cfg -O model/yolov3.cfg

Expected output:

Model loaded: YOLOv3
CUDA available, using GPU for inference

Accuracy: ★★☆☆☆
Speed: ★★★☆☆
GPU Support: No (CPU only)Used when no YOLO models are found:

Warning: YOLO weights not found. Using OpenCV's built-in HOG person detector as fallback.
Model loaded: HOG

Limitations:

Lower accuracy, especially with occlusions
Struggles with non-upright poses
More false positives

Only use HOG for testing. Download YOLO models for production.

Using Test Results for Production

Finding Optimal Settings

Test with multiple images

Capture diverse scenarios from your deployment:

# Different times of day
uv run main.py --test-image morning.jpg --save image
uv run main.py --test-image afternoon.jpg --save image
uv run main.py --test-image evening.jpg --save image

# Different distances
uv run main.py --test-image close-up.jpg --save image
uv run main.py --test-image medium-range.jpg --save image
uv run main.py --test-image far-away.jpg --save image

Tune for your use case

High-security (minimize false negatives):

uv run main.py --test-image test.jpg --save image \
  --confidence 0.4 \
  --area-threshold 800

High-precision (minimize false positives):

uv run main.py --test-image test.jpg --save image \
  --confidence 0.7 \
  --area-threshold 3000

Document your findings

Record optimal settings for your environment:

# production.cfg
[paths]
model_dir = model
output_dir = /data/surveillance

[detection]
confidence_threshold = 0.6  # Tuned from test results
person_area_threshold = 2500  # Filters distant persons
frame_skip = 15

Apply to live streams

Use tested settings with RTSP streams:

uv run main.py --config production.cfg \
  --rtsp "rtsp://camera.local/stream" \
  --save image

Advanced Testing

Batch Testing Multiple Images

#!/bin/bash
# test-batch.sh

for img in test_images/*.jpg; do
  echo "Testing: $img"
  uv run main.py --test-image "$img" --save image \
    --confidence 0.5 \
    --area-threshold 2000
  echo "---"
done

Run:

chmod +x test-batch.sh
./test-batch.sh

Comparing Threshold Ranges

#!/bin/bash
# compare-thresholds.sh

for conf in 0.3 0.4 0.5 0.6 0.7; do
  echo "Testing confidence: $conf"
  uv run main.py --test-image test.jpg --save image --confidence $conf
  mv test_result_*.jpg "result_conf_${conf}.jpg"
done

Review all result_conf_*.jpg files to compare detection results.

Testing GPU Acceleration

Verify CUDA is being used:

uv run main.py --test-image test.jpg --save image 2>&1 | grep CUDA

Expected output:

CUDA available, using GPU for inference

If you see:

CUDA not available, using CPU for inference

Install CUDA-enabled OpenCV:

uv pip install opencv-contrib-python

Image Mode

Apply tested settings to capture snapshots from live streams

Video Mode

Record MP4 clips with your tuned detection parameters

Get Started

Guides

Examples

Overview

Basic Usage

Step-by-Step Guide

Understanding the Output

Console Output Breakdown

Annotated Image

Tuning Detection Parameters

Confidence Threshold Examples

Area Threshold Examples

Troubleshooting Detection Issues

Comparing Detection Methods

YOLOv4 vs YOLOv3 vs HOG

Using Test Results for Production

Finding Optimal Settings

Advanced Testing

Batch Testing Multiple Images

Comparing Threshold Ranges

Testing GPU Acceleration

Image Mode

Video Mode

Get Started

Guides

Examples

​Overview

​Basic Usage

​Step-by-Step Guide

​Understanding the Output

​Console Output Breakdown

​Annotated Image

​Tuning Detection Parameters

​Confidence Threshold Examples

​Area Threshold Examples

​Troubleshooting Detection Issues

​Comparing Detection Methods

​YOLOv4 vs YOLOv3 vs HOG

​Using Test Results for Production

​Finding Optimal Settings

​Advanced Testing

​Batch Testing Multiple Images

​Comparing Threshold Ranges

​Testing GPU Acceleration

​Related Pages

Image Mode

Video Mode

Overview

Basic Usage

Step-by-Step Guide

Understanding the Output

Console Output Breakdown

Annotated Image

Tuning Detection Parameters

Confidence Threshold Examples

Area Threshold Examples

Troubleshooting Detection Issues

Comparing Detection Methods

YOLOv4 vs YOLOv3 vs HOG

Using Test Results for Production

Finding Optimal Settings

Advanced Testing

Batch Testing Multiple Images

Comparing Threshold Ranges

Testing GPU Acceleration

Related Pages