Panoptic Segmentation
全景分割是计算机视觉领域的一项任务,旨在结合语义分割和实例分割,以提供场景的全面理解。其目标是将图像分割成具有语义意义的部分或区域,并检测和区分这些区域内的各个对象实例。每个像素都被分配一个语义标签,而属于“事物”类的像素(如可计数的对象实例)则被赋予唯一的实例ID。
ADE20K
MasQCLIP
ADE20K val
OneFormer (DiNAT-L, single-scale, 1280x1280, COCO-Pretrain)
Cityscapes test
OneFormer (ConvNeXt-L, single-scale, Mapillary Vistas-Pretrained)
Cityscapes val
Panoptic FCN* (Swin-L, Cityscapes-fine)
COCO minival
OpenSeeD (SwinL, single-scale)
COCO panoptic
VAN-B6*
COCO test-dev
Mask DINO (single scale)
DALES
SuperCluster
Hypersim
Indian Driving Dataset
EfficientPS
KITTI-360
KITTI Panoptic Segmentation
EfficientPS
LaRS
Mask2Former (Swin-B)
Mapillary val
OneFormer (DiNAT-L, single-scale)
MUSES: MUlti-SEnsor Semantic perception dataset
NYU Depth v2
PanNuke
LKCell
Panoptic nuScenes test
(AF)2-S3Net + CenterPoint
Panoptic nuScenes val
PASTIS
Exchanger+Mask2Former
PASTIS-R
Early Fusion
S3DIS
S3DIS Area5
ScanNet
OneFormer3D
ScanNetV2
OneFormer3D
SemanticKITTI
P3Former
SUN-RGBD