Event Detection & Temporal Intelligence Expert

Need to group photos into meaningful events?
├─ Have GPS + timestamps? ──────────────────── ST-DBSCAN
│   ├─ Also need visual similarity? ────────── DeepDBSCAN (add CLIP)
│   └─ Need hierarchical events? ───────────── Multi-level cascading
│
├─ No GPS, only timestamps? ────────────────── Temporal binning
│   └─ With visual content? ─────────────────── CLIP + temporal
│
└─ Photos have faces + want groups? ─────────── Face clustering first
    └─ Then event detection per person

Core Concepts

1. ST-DBSCAN: Spatio-Temporal Clustering

The Problem: Standard clustering fails for photos—same location on different days shouldn't be grouped.

Key Insight: 100 meters apart in same hour = same event. 100 meters apart 3 days later = different events.

ST-DBSCAN Parameters:

ε_spatial:   50m (indoor) → 500m (outdoor festival) → 5km (city tour)
ε_temporal:  1hr (short event) → 8hr (day trip) → 24hr (multi-day)
min_pts:     3 (small gathering) → 10 (large event)

Algorithm: Both spatial AND temporal constraints must be satisfied:

Neighbor(p) = {q | distance(p,q) ≤ ε_spatial AND |time(p)-time(q)| ≤ ε_temporal}

→ Deep dive: references/st-dbscan-implementation.md

2. DeepDBSCAN: Adding Visual Content

Problem: Photos at same time/place can be different subjects (ceremony vs empty chairs).

Solution: Add CLIP embeddings as third dimension:

Neighbor(p) = {q | spatial_ok AND temporal_ok AND cosine_sim(clip_p, clip_q) > threshold}

eps_visual: 0.3 (similar subjects) → 0.5 (diverse event content)

3. Hierarchical Event Detection

Use case: "Paris Vacation" contains "Day 1: Louvre", "Day 2: Eiffel Tower"

Approach: Cascade ST-DBSCAN with expanding thresholds:

1. High-level (vacations): eps_spatial=50km, eps_temporal=72hr
2. Mid-level (daily): eps_spatial=5km, eps_temporal=12hr
3. Low-level (moments): eps_spatial=500m, eps_temporal=1hr

Event Significance Scoring

Goal: Birthday party > Daily commute photos

Multi-Factor Model (weights sum to 1.0):

→ Deep dive: references/event-scoring-shareability.md

Shareability Prediction

Goal: Predict which photos will be shared on social media.

High-Signal Features (2025 research):

1. Smiling faces (+0.3 base score)
2. Group photos (3+ people, +0.2)
3. Famous landmarks (+0.25)
4. Food scenes (+0.15)
5. Moderate visual complexity (0.4-0.6 optimal)
6. Recency (decays over 30 days)

Shareability Threshold: >0.6 = "Highly Shareable"

→ Deep dive: references/event-scoring-shareability.md

Life Event Detection

Automatically detect major life events using multi-modal signals:

→ Deep dive: references/place-recognition-life-events.md

Temporal Diversity for Selection

Problem: Without constraints, collage might be all vacation photos.

Method Comparison

Temporal MMR Formula

MMR(photo) = λ × quality + (1-λ) × min_temporal_distance_to_selected

• λ=0.5: Balanced
• λ=0.7: Prefer quality
• λ=0.3: Prefer diversity

→ Deep dive: references/temporal-diversity-pipeline.md

Common Anti-Patterns

Anti-Pattern: Time-Only Clustering

What it looks like: Using K-means or basic DBSCAN on timestamps only

clusters = KMeans(n_clusters=10).fit(timestamps)  # WRONG

Why it's wrong: Multi-day trips at same location get split; same-day different-location events get merged.

What to do instead: Use ST-DBSCAN with both spatial AND temporal constraints.

Anti-Pattern: Fixed Epsilon Values

What it looks like: Using same eps_spatial=100m for all events

Why it's wrong: Indoor events need 50m, city tours need 5km.

What to do instead: Adaptive thresholds based on event type detection, or hierarchical clustering with multiple scales.

Anti-Pattern: Ignoring Visual Content

What it looks like: ST-DBSCAN alone for event detection

Why it's wrong: Wedding ceremony and empty chairs setup—same time/place, completely different importance.

What to do instead: DeepDBSCAN with CLIP embeddings for content-aware clustering.

Anti-Pattern: Euclidean Distance for GPS

What it looks like:

distance = sqrt((lat2-lat1)**2 + (lon2-lon1)**2)  # WRONG

Why it's wrong: Degrees ≠ meters. 1° latitude = 111km, but 1° longitude varies by latitude.

What to do instead: Haversine formula for great-circle distance:

from geopy.distance import geodesic
distance_meters = geodesic((lat1, lon1), (lat2, lon2)).meters

Anti-Pattern: No Noise Handling

What it looks like: Forcing every photo into a cluster

Why it's wrong: Solo commute photos pollute event clusters.

What to do instead: DBSCAN naturally identifies noise (label=-1). Keep noise separate—don't force into nearest cluster.

Anti-Pattern: Shareability Without Event Context

What it looks like: Predicting shareability from photo features alone

Why it's wrong: A mediocre photo from your wedding is more shareable than a great photo from Tuesday's lunch.

What to do instead: Include event significance as feature:

features['event_significance'] = photo.event.significance_score

Quick Start: Event Detection Pipeline

from event_detection import EventDetectionPipeline

pipeline = EventDetectionPipeline()

# Process photo corpus
results = pipeline.process_photo_corpus(photos)

# Access events
for event in results['events']:
    print(f"{event.label}: {len(event.photos)} photos, significance={event.significance_score:.2f}")

# Access life events
for life_event in results['life_events']:
    print(f"{life_event.type} detected on {life_event.timestamp}")

# Select for collage with diversity
collage_photos = pipeline.select_for_collage(results, target_count=100)

Performance Targets

Python Dependencies

numpy scipy scikit-learn hdbscan geopy transformers xgboost pandas opencv-python

Integration Points

• collage-layout-expert: Pass event clusters for diversity-aware placement
• photo-content-recognition-curation-expert: Get face clusters before event detection
• color-theory-palette-harmony-expert: Use for visual diversity within events
• clip-aware-embeddings: Generate embeddings for DeepDBSCAN

References

1. ST-DBSCAN: Birant & Kut (2007), "ST-DBSCAN: An algorithm for clustering spatial-temporal data"
2. DeepDBSCAN: ISPRS 2021, "Deep Density-Based Clustering for Geo-Tagged Photos"
3. Shareability: arXiv 2025, "Predicting Social Media Engagement from Emotional and Temporal Features"
4. GeoNames/OpenStreetMap: Reverse geocoding for place recognition

Version: 2.0.0 Last Updated: November 2025