IoU metric

In the docs of iou_score you state only that predictions are of the shape (B, H, W, C) but you state nothing about the values in predictions. Are those probabilities (let’s say [0.2, 0.1, 0., 0.7] for one pixel and the case with 4 class) or the one-hot encoded vectors ([0, 0, 0, 1] for the same pixel and the same number of classes)?

Because the result is not the same. Let’s say that the ground truth in this example is: [0, 1, 0, 0]. In the first case, the intersection is 0.2, but in the second case, it is 0. And 0 is the correct intersection value.

Actually, I’m not sure how to interpret it if those are probabilities …

[qubvel]

Ok, I got it. I will add this feature.

[Tyler-D]

Agree with @ajuric . When training using softmax on multi-classes task (PASCAL VOC for example), I apply argmax to final feature map HWC to get final prediction H*W mask (each cell is the class_id ). And there is no threshold needed actually. Then I compute IOU based on the ground truth mask and prediction mask. I think it will be great if we have both kinds of IOU here.