Panoramic Semantic Segmentation

Panoramic semantic segmentation aims to automatically label every pixel in a 360° image with its corresponding semantic class, a challenging task due to image distortions and the scarcity of labeled panoramic data. Current research focuses on addressing these challenges through unsupervised domain adaptation techniques, leveraging readily available pinhole images or synthetic data, and employing transformer-based architectures or convolutional neural networks with specialized modules to handle distortions and object deformations. These advancements are significant for improving scene understanding in various applications, such as autonomous driving and virtual/augmented reality, where comprehensive 360° scene perception is crucial.