OpenPose overview in Computer Vision - Model Metrics & Evaluation

Metrics & Evaluation - OpenPose overview

Which metric matters for OpenPose and WHY

OpenPose detects body keypoints like joints in images or videos. The main metric to check how well it works is Percentage of Correct Keypoints (PCK). PCK measures how many predicted points are close enough to the true points. This matters because OpenPose needs to find exact body parts to be useful.

Confusion matrix or equivalent visualization

    True Keypoints:    O O O O O
    Predicted Points:  O O X O O

    Here, 'O' means correct keypoint detected,
    'X' means missed or wrong keypoint.

    Count of correct points (TP): 4
    Count of missed points (FN): 1
    False positives (FP) are rare since OpenPose predicts fixed points.
    

Precision vs Recall tradeoff with examples

In OpenPose, precision means how many detected points are actually correct. Recall means how many true points were found.

If precision is high but recall is low, OpenPose finds few points but they are mostly right. This is safe but misses body parts.

If recall is high but precision is low, OpenPose finds many points but some are wrong. This can confuse applications.

For example, in sports analysis, high recall is important to track all movements. In medical use, high precision is needed to avoid wrong body part detection.

What "good" vs "bad" metric values look like for OpenPose

Good: PCK above 85% means most keypoints are detected accurately. Precision and recall both above 80% show balanced and reliable detection.

Bad: PCK below 60% means many keypoints are missed or wrong. Precision or recall below 50% means the model is unreliable for real use.

Common pitfalls in OpenPose metrics

• Ignoring scale: Keypoint distance thresholds must adjust for image size, or PCK will be misleading.
• Overfitting: Model may work well on training poses but fail on new people or angles.
• Data leakage: Testing on images similar to training can inflate metrics falsely.
• Ignoring occlusions: Missing points due to blocked body parts can lower recall unfairly.

Self-check question

Your OpenPose model has 90% precision but only 40% recall. Is it good for tracking full body movement? Why or why not?

Answer: No, because it misses many true keypoints (low recall). It finds few points but mostly correct. For full body tracking, missing many points is a problem.

Key Result

Percentage of Correct Keypoints (PCK) is key to measure OpenPose accuracy in detecting body parts.