curiositech/event-detection-temporal-intelligence-expert

Event Detection & Temporal Intelligence Expert

Expert in detecting meaningful events from photo collections using spatio-temporal clustering, significance scoring, and intelligent photo selection.

DECISION POINTS

ST-DBSCAN vs DeepDBSCAN Algorithm Selection

Photo corpus analysis needed?
├─ Have GPS + timestamps?
│   ├─ Same location, different activities detected? ──── DeepDBSCAN
│   │   └─ Cost tolerance: High accuracy > speed ─────── Add CLIP embeddings
│   └─ Simple time/location grouping sufficient? ────── ST-DBSCAN
│
├─ Timestamps only (no GPS)?
│   ├─ Visual similarity important? ───────────────── Temporal + CLIP clustering  
│   └─ Pure time-based events? ────────────────────── Temporal binning
│
└─ Need hierarchical events (vacation > daily > moments)?
    └─ Multi-level ST-DBSCAN cascade ─────────────────── Expanding ε thresholds

Parameter Selection Matrix

| Event Type | ε_spatial | ε_temporal | min_pts | Use Case | |------------|-----------|------------|---------|----------| | Indoor party | 50m | 4hr | 5 | Home gatherings | | Wedding | 200m | 8hr | 8 | Venue + reception | | City tour | 5km | 12hr | 3 | Tourism, exploration | | Multi-day trip | 50km | 72hr | 10 | Vacation clustering | | Conference | 1km | 24hr | 6 | Business events |

Event Significance Threshold Decision

Computed significance score?
├─ Score ≥ 0.8? ──────── Life event candidate (birth, wedding, graduation)
├─ Score ≥ 0.6? ──────── Major memorable event  
├─ Score ≥ 0.4? ──────── Significant social gathering
├─ Score ≥ 0.2? ──────── Minor event worth keeping
└─ Score < 0.2? ──────── Daily routine, consider filtering

FAILURE MODES

Over-Clustering Syndrome

Symptoms: Every few photos become separate "events"; 50+ micro-events from one vacation Detection Rule: If >30% of events contain <5 photos AND duration <2 hours Diagnosis: ε parameters too restrictive, treating natural breaks as separate events Fix: Increase ε_temporal (2hr → 6hr) or use hierarchical clustering with larger top-level ε

Under-Clustering Collapse

Symptoms: Wedding ceremony + reception next day grouped as single event Detection Rule: If single event spans >24hr AND contains >200 photos AND location changes >5km Diagnosis: ε parameters too permissive, merging distinct occasions Fix: Reduce ε_temporal (12hr → 6hr) OR add location change detection as break condition

GPS Noise Contamination

Symptoms: Indoor event scattered across 10km radius; bathroom photos 500m from venue Detection Rule: If event location std_dev >2x expected venue size AND contains <20% outdoor photos Diagnosis: GPS drift/reflection causing false spatial spread Fix: Apply GPS smoothing filter OR increase min_pts to require more spatial consensus

Content-Blind Grouping

Symptoms: Empty venue setup photos grouped with ceremony; parking lot + wedding altar same event Detection Rule: If visual diversity within event >0.8 cosine distance AND high location precision Diagnosis: ST-DBSCAN without visual validation grouping unrelated content Fix: Switch to DeepDBSCAN with ε_visual=0.4 OR post-filter by CLIP similarity

Temporal Boundary Bleeding

Symptoms: Friday work photos grouped with Saturday family party; overnight events split at midnight Detection Rule: If event crosses date boundary AND activity types differ >0.6 semantic distance Diagnosis: Fixed temporal windows ignoring natural event boundaries Fix: Use adaptive temporal windows OR detect activity changes as natural breaks

WORKED EXAMPLES

Example 1: Wedding Event Detection

Input: 847 photos from weekend wedding, GPS enabled

photos = load_wedding_corpus("sarah_tom_wedding/")
# GPS range: Venue (40.7589, -73.9851) to Hotel (40.7505, -73.9934)  
# Time range: Fri 2pm - Sun 11am (45 hours)

Decision Process:

1. Algorithm Choice: Multiple venues + high importance → DeepDBSCAN
2. Parameter Selection: Wedding type → ε_spatial=200m, ε_temporal=8hr, min_pts=8
3. Visual Threshold: Diverse wedding activities → ε_visual=0.5

Expert vs Novice Decisions:

• Novice miss: Would use single ε_temporal=24hr, grouping rehearsal dinner with ceremony
• Expert catch: Detects natural breaks (rehearsal→ceremony→reception) using CLIP similarity drops

Results:

Event 1: Rehearsal Dinner (Fri 6pm-10pm, 34 photos)
Event 2: Getting Ready (Sat 10am-2pm, 89 photos) 
Event 3: Ceremony (Sat 2pm-4pm, 156 photos)
Event 4: Reception (Sat 5pm-11pm, 203 photos)
Noise: Travel/hotel photos (47 photos)

Example 2: Parameter Trade-off Analysis

Scenario: 10,000 photo family corpus, computational budget constraints

Trade-off Decision:

• ST-DBSCAN: 2.1sec processing, 85% event accuracy, 12 false merges
• DeepDBSCAN: 47sec processing, 94% event accuracy, 3 false merges

Decision Factors:

• Real-time requirement? → ST-DBSCAN
• Batch processing + accuracy critical? → DeepDBSCAN
• Hybrid: ST-DBSCAN with DeepDBSCAN refinement on high-significance events

Expert Insight: Cost/accuracy inflection point at ~5,000 photos where CLIP embedding overhead becomes worthwhile.

QUALITY GATES

Event detection task complete when ALL conditions met:

• [ ] Every photo assigned to event cluster OR explicitly marked as noise (-1 label)
• [ ] No event spans >48 hours without explicit multi-day justification
• [ ] No single-photo events (unless min_pts=1 explicitly chosen for outlier detection)
• [ ] Event significance scores computed and >0.8 flagged for life event analysis
• [ ] GPS coordinate outliers (>3 std_dev from event centroid) investigated for noise
• [ ] Visual diversity within events <0.7 cosine distance (if using DeepDBSCAN)
• [ ] Temporal gaps >6 hours within events have activity continuity justification
• [ ] Parameter choices documented with rationale (indoor/outdoor, event type)
• [ ] Processing time <5sec per 1000 photos (performance requirement)
• [ ] Event hierarchy validated if multi-level clustering applied

NOT-FOR BOUNDARIES

Do NOT use this skill for:

• Individual photo quality assessment → Use photo-composition-critic instead
• Color scheme analysis for events → Use color-theory-palette-harmony-expert instead
• Face recognition/clustering → Use photo-content-recognition-curation-expert first, then apply event detection
• EXIF timestamp extraction → Basic file parsing, not event intelligence
• Single photo context → This skill requires photo collections (>10 photos minimum)
• Real-time photo stream processing → Designed for batch corpus analysis
• Geographic route planning → Use mapping services; this extracts events from completed trips
• Social graph analysis → This handles temporal/spatial clustering, not relationship mapping

Delegation patterns:

• For photo aesthetic quality within events → photo-composition-critic
• For face-based event grouping → First photo-content-recognition-curation-expert, then this skill
• For collage layout of detected events → collage-layout-expert
• For color harmony across event photos → color-theory-palette-harmony-expert