Metrics for segmentation

[talhaanwarch]

Do, we have metrics for segmentation tasks?
I checked here, it only classification and regression

[SkafteNicki]

Hi @talhaanwarch,
Sorry for not getting back to you sooner. Which metrics are usually used for segmentation?

[paudom]

To mention the most common:

• DICE Score
• mIoU (mean Intersection over Union)

[Sarmadfismael]

can we use these metrics for sematic segmentation ??

[Borda]

@talhaanwarch @paudom @Sarmadfismael — TorchMetrics now has a dedicated segmentation module! Here's what's available (v1.9.0):

torchmetrics.segmentation:

Metric	Class
Dice Score	DiceScore
Mean IoU	MeanIoU
Hausdorff Distance	HausdorffDistance
Generalized Dice	GeneralizedDiceScore
Panoptic Quality	PanopticQuality

Quick example — semantic segmentation:

from torchmetrics.segmentation import DiceScore, MeanIoU

dice = DiceScore(num_classes=21, average="macro", input_format="index")
miou = MeanIoU(num_classes=21, input_format="index")

# preds: (N, H, W) — predicted class indices
# target: (N, H, W) — ground truth
dice_score = dice(preds, target)
miou_score = miou(preds, target)

With Lightning:

from torchmetrics import MetricCollection
from torchmetrics.segmentation import DiceScore, MeanIoU

seg_metrics = MetricCollection({
    "dice": DiceScore(num_classes=21, average="macro", input_format="index"),
    "miou": MeanIoU(num_classes=21, input_format="index"),
})
self.val_metrics = seg_metrics.clone(prefix="val/")

def validation_step(self, batch, batch_idx):
    preds = self(batch["image"]).argmax(dim=1)
    self.log_dict(self.val_metrics(preds, batch["mask"]))

Docs: DiceScore | MeanIoU