Semantic Segmentation

Image segmentation can be divided into three main types: semantic segmentation (assigning a class to each pixel), instance segmentation (distinguishing different individuals of the same class), and panoptic segmentation (combining semantic segmentation and instance segmentation).

In navigation systems, misclassifying obstacles as traversable areas (false traversable) is the most fatal type of error.

In safety-critical navigation scenarios, metrics more meaningful than mIoU include obstacle class recall and False Traversable Rate.

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.

Semantic segmentation (such as U-Net) places higher demands on data annotation quality and evaluation metrics (mIoU).

Semantic segmentation requires assigning a category label to every pixel in an image, and manually annotating a single complex image can take tens of minutes.

Image segmentation is commonly achieved using semantic segmentation models in deep learning such as U-Net or Mask R-CNN, which determine pixel by pixel which regions belong to the bird's main body.

Layer separation technology integrates three core modules: semantic segmentation, depth estimation, and motion prediction.