run2_template-matching-tuning

Template Matching Parameter Tuning for Pixel-Art Games

Tool

python3 /root/.claude/skills/object_counter/scripts/count_objects.py \
    --tool count \
    --input_image <frame.png> \
    --object_image <template.png> \
    --threshold <float> \
    --dedup_min_dist <float>

Parameter Tuning Strategy

Threshold

• 0.9 (default): Strict matching. Works when objects appear identical to template.
• 0.75-0.85: Better for pixel-art games where objects have animation frames (e.g., spinning coins show different faces).
• Below 0.7: Too many false positives from similar-colored regions.

dedup_min_dist

• 3 (default): Works for very small templates but can double-count larger objects.
• 5-10: Better for game sprites (~16-32px wide). Prevents overlapping detections.

Recommended Parameters for Super Mario

Object	Threshold	dedup_min_dist	Rationale
Coin	0.75	5	Coins have multiple animation frames; lower threshold catches rotated coins
Enemy	0.9	5	Goombas are distinctive and consistent in appearance
Turtle	0.9	5	Koopa turtles are distinctive; 0.9 avoids false positives from terrain

Tuning Process

1. Start with threshold=0.9, dedup=3 (recommended defaults)
2. Visually inspect each frame to count expected objects
3. Sweep thresholds (0.9, 0.85, 0.8, 0.75, 0.7) while checking for:
   • False negatives (objects missed → lower threshold)
   • False positives (non-objects counted → raise threshold or dedup_min_dist)
4. Increase dedup_min_dist if duplicate detections appear at lower thresholds
5. Different object types may need different optimal thresholds

Key Insight

The HUD coin icon (top of screen) is very small and typically doesn't match the template at normal thresholds. Game coins in the play area are larger and match well.