If the IoU threshold becomes stricter, fewer predicted boxes may be counted as true positives.

IoU measures the overlap between the predicted bounding box and the ground-truth bounding box.

mAP is the same thing as accuracy, so it only measures the proportion of correct predictions.

AP summarises the precision–recall performance of a detector, rather than just checking one fixed confidence threshold.

Confidence score helps decide whether a predicted bounding box is kept or reported by the model.

A prediction with high confidence is always counted as a true positive, even if the predicted box does not overlap well with the ground-truth box.

Image segmentation assigns labels at pixel level.

Image segmentation only gives one label for the whole image.

Segmentation can be useful when estimating whether a wall is 30%, 50%, or 80% painted.

Image segmentation is usually less fine-grained than image classification.

A segmentation model can produce a mask showing the region of interest.

If training loss decreases but validation performance becomes worse, this may indicate overfitting.

The model’s performance on validation or test data is guaranteed to improve.

A higher epoch number always means the final model is better.

The training loss may generally decrease during training, but it is not guaranteed to decrease perfectly at every single step.

The risk of overfitting may increase, especially if the model keeps fitting patterns that are specific to the training data.

The training process will usually take longer to run.