Segmentation evaluation (IoU, Dice) in Computer Vision - Model Metrics & Evaluation

Metrics & Evaluation - Segmentation evaluation (IoU, Dice)

Which metric matters for segmentation and WHY

For segmentation tasks, we want to see how well the predicted area matches the true area. Two main metrics help us do this:

• IoU (Intersection over Union): It measures the overlap between the predicted and true segments divided by their combined area. It tells us how much the prediction and truth agree.
• Dice coefficient: It is similar but gives more weight to the overlap. It is twice the overlap divided by the total size of both areas.

Both metrics range from 0 to 1, where 1 means perfect match. They help us understand how accurate the segmentation is in a clear way.

Confusion matrix for segmentation (pixel-wise)

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

      IoU = TP / (TP + FP + FN)
      Dice = 2 * TP / (2 * TP + FP + FN)
    

Here, each pixel is counted as positive if it belongs to the object and negative if it does not. TP means pixels correctly predicted as object, FP means pixels wrongly predicted as object, and FN means pixels missed.

Tradeoff: IoU vs Dice with examples

Both IoU and Dice measure overlap but differ slightly:

• IoU is stricter. If the prediction is smaller or larger than the true area, IoU drops more.
• Dice is more forgiving and often higher than IoU for the same prediction.

Example: If your model predicts a tumor area in a scan:

• IoU tells you exactly how much the predicted tumor overlaps the real tumor.
• Dice gives a smoother score that can be easier to optimize during training.

Choosing between them depends on your task needs. Dice is popular in medical imaging, IoU in general object detection.

What good vs bad metric values look like

• Good: IoU or Dice above 0.8 means the predicted segment closely matches the true segment.
• Moderate: Scores between 0.5 and 0.8 show partial overlap but room for improvement.
• Bad: Scores below 0.5 mean poor overlap, predictions are often missing or wrong.

For example, a Dice score of 0.9 means 90% overlap, which is excellent for tasks like tumor segmentation.

Common pitfalls in segmentation metrics

• Ignoring class imbalance: If the object is very small, high Dice or IoU can be hard to achieve and accuracy can be misleading.
• Overfitting: Very high scores on training data but low on new images mean the model memorized rather than learned.
• Boundary errors: Small shifts in edges can reduce IoU a lot even if the shape looks visually good.
• Data leakage: Testing on images seen during training inflates metrics falsely.

Self-check question

Your model has an IoU of 0.95 on training images but only 0.4 on new images. Is it good for production? Why or why not?

Answer: No, this shows overfitting. The model performs well on known data but poorly on new data, so it will not generalize well in real use.

Key Result

IoU and Dice measure how well predicted segments overlap true segments; high values (close to 1) mean better segmentation quality.