Interactive error visualization in image confusion matrix
By default, the error mode shows all labeled pixels in an image. In order to gain deeper insight into model performance, it is possible to limit the visualization to specific type of errors in a confusion matrix.
Hovering the mouse over the image confusion matrix, we visualize only the corresponding subset of pixels.
Using the French fries example from previous page, we visualize only examples labeled a potato skin:
When we move the mouse to the entry potato skin - rotten (the value of 27), we can quickly identify the part of the stroke where the current models mislabels skin as rot defect.
Important: Please note, that the green/red pixel visualization corresponds to the current model. There may be several reasons why we observe incorrect labeling in the left part of the stroke:
- Our labeling may not be precise and the error suggests which part is incorrectly labeled. In this most common case, removing the potato skin on the left and the right end of the stroke would be advisable to improve the trainng labels. Leaving incorrect labels confuses machine learning models.
- If we observe model errors when inspecting a test image (unused in model building), it may be that the areas highlighted as errors are, in fact, valid examples not represented in the training set. In such a case, it is advisable to include similar material examples to our training set.