Lesson 022: Heuristic Month-Fit Scoring Without Text Metadata

The Lesson

When images lack text metadata (titles, descriptions, captions), month or season suitability can still be approximated from visual features alone — color temperature, brightness, contrast, and content flags. The signal is coarse (3-4 seasonal buckets, not 13 distinct months) but sufficient to prevent the worst mismatches and break ties in an optimizer.

Context

A calendar optimization needed to assign 13 images to 13 months (December 2026 through December 2027). The 12,217 vote-pool images had visual features (brightness, contrast, saturation, dominant colors, content flags for Earth/Moon/crew/spacecraft) and CLIP embeddings, but zero text metadata — no titles, no descriptions, no captions. Month-fit scoring had to work entirely from visual signals.

What Happened

1. Defined target profiles for each of 13 months: brightness level, color warmth, visual drama (contrast × saturation), and content-flag bonuses (e.g., Earth bonus for April/Earth Day, spacecraft bonus for launch month).
2. Extracted color temperature from dominant color JSON by converting RGB to HSV hue: warm hues (reds, oranges, 0-60° and 300-360°) map to warmth ≈ 1.0, cool hues (blues, cyans, 180-260°) map to warmth ≈ 0.0. Low-saturation images (common in space photos) get neutral warmth (0.5) to avoid penalizing dark space scenes.
3. Scored each image against each month as: 1 - weighted_distance(image_features, month_profile), with content-flag bonuses added. Weights: 30% brightness distance, 35% warmth distance, 35% drama distance.
4. Observed that most space images cluster in the low-brightness, low-saturation, cool-temperature region. Month-fit scores were compressed — most images scored similarly for most months (range ~0.5-0.85). The signal discriminates at the margins, not across the full range.
5. Used month-fit at 15% weight in the MMR greedy selector — enough to break ties but not enough to override strong preference signals.
6. The Hungarian algorithm for month assignment showed that even coarse month-fit signals produce meaningfully different assignments than random: Method D (month-first) achieved total month-fit of 10.5 vs. Method A's 8.5.

Key Insights

• Coarse signal beats no signal. A heuristic that distinguishes "warm bright summer" from "dark dramatic winter" in 3-4 buckets is enough to prevent bad assignments (bright warm Earth shot on a November page). It doesn't need to distinguish March from April.

• Space imagery has compressed visual feature distributions. Most mission photos are dark, low-saturation, and cool-toned. The month-fit heuristic works best on the minority of images that have distinctive visual properties (bright Earth views, warm-toned crew shots). For the majority, month-fit is near-neutral — and that's fine.

• Weight the signal by its reliability. At 15% of the objective, month-fit breaks ties between otherwise-equal images but can't override a strong preference gap. This matches the signal's quality — it's a heuristic, not a measurement.

• Content flags provide the strongest month discrimination. An image with has_earth_flag=true gets a bonus for April (Earth Day). An image with has_spacecraft_flag=true gets a bonus for December 2026 (launch month). These categorical signals are more informative than continuous visual features for month assignment.

• The assignment algorithm compensates for noisy scores. The Hungarian algorithm finds the globally optimal assignment even when individual month-fit scores are noisy. Small differences in scores still produce better assignments than random allocation.

Applicability

This lesson applies when you need to assign items to slots based on "fit" but lack rich metadata:

• Assigning stock photos to seasonal marketing campaigns using only visual features
• Ordering songs in a playlist by mood using only audio features (tempo, key, energy)
• Arranging art in a gallery by visual flow using only color/composition analysis

Does NOT apply when rich text metadata is available — in that case, NLP-based semantic matching will outperform visual heuristics.