MediaPipe Pose in Computer Vision - Model Metrics & Evaluation

Metrics & Evaluation - MediaPipe Pose

Which metric matters for MediaPipe Pose and WHY

For MediaPipe Pose, the key metric is Mean Average Precision (mAP) or Percentage of Correct Keypoints (PCK). These metrics measure how accurately the model detects body landmarks compared to the true positions.

We care about these because the goal is to find exact points on the body like elbows or knees. If the points are off, the pose estimation is wrong. So, accuracy in locating these points is critical.

Also, inference speed matters because pose detection often runs live on video. A slow model makes the experience laggy.

Confusion matrix or equivalent visualization

Pose estimation does not use a classic confusion matrix because it predicts many points per image. Instead, we use a distance threshold to decide if a predicted keypoint is correct.

    True Keypoint Positions:    (x1, y1), (x2, y2), ...
    Predicted Keypoint Positions: (x1', y1'), (x2', y2'), ...

    For each keypoint:
      If distance(predicted, true) < threshold: count as True Positive (TP)
      Else: False Positive (FP) or False Negative (FN) depending on missing points

    Total keypoints = TP + FP + FN
    

This helps calculate Precision, Recall, and F1 score for keypoint detection.

Precision vs Recall tradeoff with examples

Precision means how many detected keypoints are actually correct. High precision means few false points.

Recall means how many true keypoints the model found. High recall means few missed points.

For example, in a fitness app, missing a keypoint (low recall) can cause wrong exercise feedback. So recall is very important.

But if the model detects many wrong points (low precision), the app might confuse the user. So precision also matters.

We balance precision and recall to get a good overall F1 score, ensuring the model finds most points and keeps them accurate.

What "good" vs "bad" metric values look like for MediaPipe Pose

Good values:

• Precision > 0.85 (85%) - Most detected points are correct
• Recall > 0.85 (85%) - Most true points are found
• F1 score > 0.85 - Balanced and accurate detection
• Inference speed < 30 ms per frame - Real-time performance

Bad values:

• Precision < 0.6 - Many false points confuse the system
• Recall < 0.6 - Many true points missed, poor pose estimation
• F1 score < 0.6 - Overall poor detection quality
• Inference speed > 100 ms per frame - Laggy and unusable live

Common pitfalls in MediaPipe Pose metrics

• Ignoring speed: A model with high accuracy but slow speed is not practical for live pose detection.
• Overfitting: Model performs well on training videos but poorly on new people or backgrounds.
• Data leakage: Testing on the same videos used for training inflates accuracy falsely.
• Using accuracy alone: Accuracy can be misleading because many keypoints are easy to detect; focus on precision, recall, and F1.
• Threshold choice: Setting the distance threshold too loose or tight changes metric results unfairly.

Self-check question

Your MediaPipe Pose model has 98% accuracy but only 12% recall on keypoints. Is it good for production? Why or why not?

Answer: No, it is not good. The very low recall means the model misses most true keypoints, so it fails to detect the full pose. High accuracy alone is misleading because many keypoints might be absent or ignored. For pose estimation, recall is critical to find all body points.

Key Result

For MediaPipe Pose, balanced high precision and recall (above 85%) with fast inference speed are key to good pose detection.