IoU

The Dice coefficient is defined as twice the intersection area divided by the sum of the areas of the two regions, and is highly mathematically correlated with IoU, but Dice is more robust to class imbalance, and when the target region is much smaller than the background region, Dice is usually more suitable than IoU as the primary evaluation metric.

Under the configuration of image size 300×300 and batch size 16, UNet++ achieved the highest IoU of 0.6719 and DICE of 0.7869, performing best among all compared models.

IoU (intersection over union) is the area of intersection between the predicted box and the ground-truth box divided by the area of their union, ranging from 0 to 1, with values closer to 1 indicating a more accurate box.

IoU is the most basic metric in grounding tasks, computed as the intersection area of the predicted box and the ground-truth box divided by the union area, with values from 0 to 1.

In dense grounding scenarios, common failure modes include missing small objects, merging multiple objects into one box, and coordinate drift causing a drop in IoU.

mAP is a metric averaged over multiple IoU thresholds or multiple categories, and better reflects a model's overall performance under different precision requirements.

When evaluating visual grounding, an IoU (intersection over union) threshold is generally used to determine whether a predicted box is a hit.

IoU (Intersection over Union) equals the intersection area of the predicted region and the annotated region divided by the union area, with a range between 0 and 1.

After adding 1,500 real frames from the GOOSE dataset, sky IoU improved by 7 to 17 points, vegetation IoU improved by 29 to 37 points, and real-weather mean IoU improved by 29 to 35 points across four architectures.

After adding physics-based adverse weather augmentation to a semantic segmentation model, the IoU for the real-weather "sky" category decreased across every architecture, with drops ranging from 1 to 13 percentage points.