Face recognition concept in Computer Vision - Model Metrics & Evaluation

Metrics & Evaluation - Face recognition concept

Which metric matters for Face Recognition and WHY

In face recognition, the key metrics are Precision and Recall. Precision tells us how many faces identified as a person are actually correct. Recall tells us how many faces of that person the system found out of all the faces it should find. Both matter because we want to avoid mistakes like wrongly recognizing someone (low precision) and missing a person's face (low recall). The F1 score balances these two, giving a single number to check overall performance.

Confusion Matrix for Face Recognition

      | Predicted Positive | Predicted Negative |
      |--------------------|--------------------|
      | True Positive (TP)  | False Negative (FN) |
      | False Positive (FP) | True Negative (TN)  |

      Example:
      Suppose we have 100 face images of a person.
      - TP = 85 (correctly recognized faces)
      - FN = 15 (faces missed)
      - FP = 10 (wrongly recognized as this person)
      - TN = 200 (correctly identified as not this person)

      Total samples = TP + FP + TN + FN = 85 + 10 + 200 + 15 = 310
    

Precision vs Recall Tradeoff in Face Recognition

Imagine a security system that unlocks doors using face recognition.

• High Precision: The system rarely lets in the wrong person. This is important to keep the place safe.
• High Recall: The system recognizes almost all authorized people, so they don't get locked out.

If the system focuses only on precision, it might miss some authorized people (low recall). If it focuses only on recall, it might let in some wrong people (low precision). We must find a good balance depending on what is more important.

What Good vs Bad Metrics Look Like for Face Recognition

• Good: Precision and Recall both above 90%. This means the system correctly recognizes most faces and rarely makes mistakes.
• Bad: Precision below 70% means many false recognitions. Recall below 70% means many missed faces. Either case makes the system unreliable.

Common Metrics Pitfalls in Face Recognition

• Accuracy Paradox: If most images are of different people, a model that always says "not this person" can have high accuracy but be useless.
• Data Leakage: If the same face images appear in training and testing, metrics will look better than reality.
• Overfitting: Very high training metrics but poor test metrics mean the model memorizes faces instead of learning to recognize new ones.

Self Check: Your model has 98% accuracy but 12% recall on recognizing a person's face. Is it good?

No, it is not good. The high accuracy is misleading because most images are not that person's face, so the model guesses "not this person" most of the time. The very low recall means it misses almost all faces of that person, which defeats the purpose of face recognition.

Key Result

Precision and recall are key for face recognition; high accuracy alone can be misleading.