mIoU
A widely used evaluation metric in semantic segmentation tasks. It is computed as the mean ratio of the intersection to the union of predicted and ground-truth regions across all classes; higher values indicate greater segmentation accuracy.
Timeline (last 90 days)
mIoU (mean intersection over union) is the most commonly used evaluation metric in temporal grounding tasks, computed by averaging the intersection over union of each sample's predicted interval and ground-truth interval.
mIoU is insensitive to class imbalance; even if rare but high-risk categories have extremely low IoU, the drag on overall mIoU is limited.
Semantic segmentation (such as U-Net) places higher demands on data annotation quality and evaluation metrics (mIoU).
In a semantic segmentation task, the baseline model UNet achieved an original mIoU score of 0.817.
Compared to the original weight averaging, the final solution improved mean mIoU by approximately three points, but certain stacked combinations actually decreased the score, meaning more is not always better.
Video temporal grounding refers to precisely locating the corresponding time segment in an unedited video given a natural language query, with representative works including 2D-TAN, UniVTG, TimeChat, VTimeLLM, and Moment-DETR.
All Facts (6)
mIoU (mean intersection over union) is the most commonly used evaluation metric in temporal grounding tasks, computed by averaging the intersection over union of each sample's predicted interval and ground-truth interval.
50%UnverifiedmIoU is insensitive to class imbalance; even if rare but high-risk categories have extremely low IoU, the drag on overall mIoU is limited.
50%UnverifiedSemantic segmentation (such as U-Net) places higher demands on data annotation quality and evaluation metrics (mIoU).
50%UnverifiedIn a semantic segmentation task, the baseline model UNet achieved an original mIoU score of 0.817.
50%UnverifiedCompared to the original weight averaging, the final solution improved mean mIoU by approximately three points, but certain stacked combinations actually decreased the score, meaning more is not always better.
50%UnverifiedVideo temporal grounding refers to precisely locating the corresponding time segment in an unedited video given a natural language query, with representative works including 2D-TAN, UniVTG, TimeChat, VTimeLLM, and Moment-DETR.
50%