Why features identify distinctive points in Computer Vision - Why Metrics Matter

Metrics & Evaluation - Why features identify distinctive points

Which metric matters for this concept and WHY

In computer vision, when identifying distinctive points using features, repeatability and matching accuracy are key metrics. Repeatability measures how often the same points are detected under different views or lighting. Matching accuracy shows how well features help match points between images. These metrics matter because distinctive points should be stable and unique to help tasks like object recognition or 3D reconstruction.

Confusion matrix or equivalent visualization (ASCII)

    Matching Results Confusion Matrix:

          | Matched Correctly | Matched Incorrectly |
    ------|-------------------|--------------------|
    True  |        TP         |         FN         |
    Points|                   |                    |
    ------|-------------------|--------------------|
    False |        FP         |         TN         |
    Matches|                  |                    |

    TP = Correct matches of distinctive points
    FP = Incorrect matches (wrong points matched)
    FN = Missed matches (distinctive points not matched)
    TN = Correctly identified non-matches
    

Precision vs Recall tradeoff with concrete examples

Precision means how many matched points are actually correct. High precision means few wrong matches.

Recall means how many true distinctive points are found and matched. High recall means few missed points.

Example: For a robot navigating a room, high recall is important to find enough points to understand the scene. For face recognition, high precision is important to avoid wrong matches that cause errors.

What "good" vs "bad" metric values look like for this use case

Good: Repeatability above 80%, precision and recall above 75%. This means features reliably find the same points and match them correctly.

Bad: Repeatability below 50%, precision or recall below 50%. This means features miss many points or match wrongly, making them unreliable.

Metrics pitfalls

• Overfitting: Features tuned too much on one dataset may not work well on new images.
• Data leakage: Using test images during feature design inflates metrics falsely.
• Ignoring viewpoint changes: Features that fail under rotation or scale changes have low repeatability.
• Accuracy paradox: High accuracy but low recall means many points are missed, hurting performance.

Self-check question

Your feature detector has 90% precision but only 30% recall on distinctive points. Is it good for matching in a changing environment? Why or why not?

Answer: No, because while most matches are correct (high precision), it misses many true points (low recall). This means it may fail to find enough points to match when the scene changes, reducing reliability.

Key Result

Repeatability, precision, and recall are key to measure how well features identify distinctive points reliably and uniquely.