Fix Division by Zero and Refine Per-Class Metric Calculation for Image Segmentation

[mansipatil12345]

Problem:
When calculating per-class metrics (accuracy, IoU) for image segmentation, there are two key issues that need to be addressed:

1. Division by Zero:
   In cases where a class does not appear in the ground truth for a given batch (i.e., no pixels are associated with that class), calculating per-class metrics can lead to a division by zero error. This issue could result in runtime crashes or undefined behavior.

2. Inconsistent Per-Class Metric Calculation:
   The current approach does not properly handle edge cases for rare or missing classes.
   Specifically:
   a. If a class has no samples (e.g., 0 pixels in ground truth), the metric may give misleading results.
   b. The code doesn't gracefully handle cases where certain classes are absent, which could lead to inaccurate global metrics or distortions in the evaluation process.

Proposed Fix:
Fix Division by Zero:
Ensure that when a class has no ground truth pixels, the metric computation for that class returns NaN (Not a Number) instead of causing a division by zero error. This can be done by adding a check before dividing to ensure the denominator is not zero.

Calculate accuracy per class, ensuring no division by zero occurs.

If total_per_class is greater than 0, compute accuracy as correct_per_class / total_per_class.

If total_per_class is 0 (i.e., no pixels for that class), assign NaN to avoid division by zero.

accuracy_per_class = np.where(total_per_class > 0, correct_per_class / total_per_class, np.nan)

Refine Per-Class Metric Calculation:
Improve the per-class accuracy and IoU computations by ensuring that missing classes (i.e., classes with no ground truth pixels in a batch) are handled correctly. When calculating metrics for each class, check if there are ground truth pixels for the class and avoid calculating metrics when no samples are present. This will prevent distorted values and ensure accurate evaluation.

Example of the refinement:

Ensure no calculation for classes with no samples in the ground truth

correct_per_class[class_total == 0, class_idx] = np.nan
total_per_class[class_total == 0, class_idx] = np.nan

Impact on the Project:

a. Reliability: This fix will prevent runtime errors due to division by zero and will ensure more accurate metric calculations, especially in unbalanced datasets or cases where certain classes are absent from ground truth data.
b. Usability: With the fix, the evaluation process will become more robust, providing meaningful results even when certain classes are missing.
c. Quality: The overall quality of the project’s metrics and evaluation will be improved, reducing the chance of misleading or false evaluations, especially in edge cases.

Example Code:
Here’s an example snippet showing how the current implementation may cause issues and how to apply the proposed fix:

Current Code (with potential division by zero issue):
accuracy_per_class = correct_per_class / total_per_class
Proposed Fix:
accuracy_per_class = np.where(total_per_class > 0, correct_per_class / total_per_class, np.nan)

[dpascualhe]

Could you please reformat the issue description? I'm afraid that it is not readable as it is. Also, we noticed the usage of AI in writing the description and the solution proposed. While this is not explicitly against our guidelines, make sure to properly understand the codebase and problem at hand before you start working in the PR.

[mansipatil12345]

okay , sorry sir for the inconvenience