smartdownsample 1.6.1

Smart image downsampling for image classification datasets

smartdownsample

Fast, simple image downsampling that just works

SmartDownsample samples diverse images from large collections in seconds, not hours. One simple function that works equally fast whether you're sampling 100 or 23,000 images from 24,000.

Installation

pip install smartdownsample

Features

• ⚡ Always fast - Seconds for any selection ratio
• 🎯 Smart bucketing - Better than random, faster than optimal algorithms
• 📊 Scales linearly - 24k images? No problem
• 🔧 Dead simple - One function, always works
• 🎲 Reproducible - Set seed for consistent results
• ⚖️ Honest trade-offs - Speed over perfection, good enough for most use cases

Usage

from smartdownsample import sample_diverse

# Sample 100 diverse images from 24,000 - takes seconds
selected = sample_diverse(
    image_paths=my_24k_images,
    target_count=100
)

# Sample 23,000 images from 24,000 - also takes seconds!
selected = sample_diverse(
    image_paths=my_24k_images,
    target_count=23000
)

# It's that simple.
print(f"Sampled {len(selected)} diverse images")

Visualization

Visualize selection patterns to understand algorithm behavior:

from smartdownsample import sample_diverse_with_stats
from smartdownsample import plot_bucket_distribution, plot_hash_similarity_scatter

# Get selection results + visualization data
selected, viz_data = sample_diverse_with_stats(my_images, target_count=1000)

# 1. Bucket distribution chart - shows kept vs excluded per similarity bucket
plot_bucket_distribution(viz_data['bucket_stats'], save_path="buckets.png")

# 2. Interactive scatter plot - shows visual similarity with selection status
plot_hash_similarity_scatter(viz_data, save_path="scatter.html")

Visualization is included by default:

pip install smartdownsample  # Includes matplotlib + plotly

How It Works

Simple "trim from top" algorithm that maximizes diversity while being blazing fast:

1. Hash Images (Fast)

Image → 64-bit fingerprint in ~0.01 seconds
Uses DHash with 4 parallel workers

2. Group Into Buckets (O(n))

Use first 4 hash bits to create ~16 visual groups:
Bucket A: [landscape1.jpg, landscape2.jpg, ...]     # 45 images
Bucket B: [portrait1.jpg, portrait2.jpg, ...]       # 12 images  
Bucket C: [closeup1.jpg, closeup2.jpg, ...]         # 890 images

3. Trim from Top (Ultra Fast)

Sort buckets by size (largest first)
Natural sort images within each bucket (cam01/IMG_1.jpg → cam01/IMG_10.jpg → cam02/IMG_1.jpg)
Keep ALL small buckets intact
Trim only from largest buckets using stride sampling

Example: Want 500 from 1,390 camera trap images
• 50 small buckets (1 each): Keep all = 50 images ✓
• 30 medium buckets (5 each): Keep all = 150 images ✓  
• 19 large buckets (10 each): Keep all = 190 images ✓
• 1 huge bucket (1000): Keep every 9th with camera/folder respect = 110 images ✓
Total: 500 images with maximum diversity + camera location preservation

Why It's Fast

Algorithm advantages:

• ✅ O(n) complexity - Just sort buckets once
• ✅ Stride sampling - Array slicing, not random selection
• ✅ No complex math - Simple bucket trimming
• ✅ Maximum diversity - Small buckets always preserved
• ✅ Smart folder ordering - Natural sorting preserves camera/folder structure

What you get:

• ✅ Fastest possible while maintaining quality
• ✅ Preserves rare/unique images (small buckets)
• ✅ Even sampling across camera locations and time sequences
• ✅ Natural file ordering (IMG_1.jpg → IMG_2.jpg → IMG_10.jpg)

Result: Optimal speed + maximum diversity preservation + smart camera trap ordering.

Algorithm Comparison

Approach	Speed	Diversity	Camera/Folder Aware	Use Case
Random sampling	Fastest	Poor	No	Quick tests only
smartdownsample	Ultra Fast	Excellent	Yes	Camera trap data
Complex diversity	Very Slow	Perfect	No	Research only

Real Example: 24,000 camera trap images → 1,000 selected

• Random: 1 second, poor diversity, ignores folder structure
• smartdownsample: 20 seconds, excellent diversity + respects camera locations
• Complex: 2+ hours, mathematically perfect but breaks folder grouping

Sweet spot: Maximum diversity preservation with camera-aware sampling in minimal time.

Performance

Task	Time
100 from 1,000	<5 sec
900 from 1,000	<5 sec
1,000 from 24,000	~30 sec
23,000 from 24,000	~30 sec
Any ratio	Fast ✓

Parameters

Parameter	Default	Description
image_paths	Required	List of image file paths (str or Path objects)
target_count	Required	Exact number of images to select
n_workers	4	Number of parallel workers (4 is optimal)
hash_size	8	Hash size (8 is fast and good enough)
random_seed	42	Random seed for reproducible results
show_progress	True	Whether to display progress bars

Why It's Fast

• Fixed algorithm - No switching between methods
• Simple hashing - DHash is faster than PHash
• Smart bucketing - O(n) grouping instead of O(n²) comparisons
• Natural sorting - Built-in natsort handles camera trap folder structures efficiently
• Parallel processing - But capped at 4 workers (diminishing returns above that)

License

MIT License – see LICENSE file.