DensePASS

DensePASS - 一种新的密集注释数据集，用于跨域条件下的全景分割，专门用于研究针孔到全景的传输，并附有从 Cityscapes 获得的针孔相机训练示例。 DensePASS 涵盖标记和未标记的 360 度图像，标记数据包含 19 个类别，这些类别明确符合源域（即针孔）数据中可用的类别。DensePASS is a novel densely annotated dataset for panoptic segmentation under cross-domain settings, specifically designed for studying pinhole-to-panoptic transfer, accompanied by pinhole camera training examples sourced from Cityscapes. DensePASS covers both labeled and unlabeled 360-degree images, and its labeled data contains 19 categories that explicitly match the categories available in the source domain (i.e., pinhole) data.

搜集汇总

背景与挑战

背景概述

DensePASS是一个用于跨域全景分割研究的密集注释数据集，专门处理从针孔到全景图像的传输任务。它包含19个类别的标记数据，这些类别与源域Cityscapes数据集保持一致，于2021年由Karlsruhe Institute of Technology发布。