Scale Confusion & Scale Drift

Issue: Model Fails to Distinguish Objects of Different Scales; Confuses Small/Large Versions

Frequency: Common

Symptoms

• High confidence on wrong scale version
• Confuses toy/miniature objects with real objects
• Fails at “is this thing big enough?”
• Scale-variant accuracy variance >30%

Root Cause Object recognition relies on visual patterns that are often scale-invariant (texture, color). Models trained on fixed resolutions struggle with extreme scale changes. Training data typically lacks diversity in object scales; datasets over-represent “typical” sizes.

Example

Scenario: Warehouse management
Image: Tiny model car (1cm) vs. actual car
Model: Treats both identically; assigns same class probability
Expected: Distinguish scale; apply size-dependent logic
Impact: Gripper grasps miniature; collision detection fails

Key Statistics

• Scale variance in training: 2-3x
• Accuracy at 10x scale: drops 40-50%
• Extreme scales (>100x): <30% accuracy

Mitigation Strategies

1. Multi-Scale Training: Augment with diverse object sizes (0.1x-10x)
2. Explicit Size Regression: Add auxiliary head predicting object size
3. Context Cues: Use objects of known size (reference objects) to infer scale
4. Rejection Threshold: Reject predictions if scale is atypical for object class

Metrics

• Accuracy vs. scale ratio (actual:expected)
• Scale regression error (predicted size / actual size)

Alerts

• Scale prediction error >50% → P2
• Confidence-calibrated for scale → Track separately

References

• Multi-Scale Feature Learning
• Scale-Aware Object Detection