search

lxtGH / OMG-Seg

OMG-LLaVA and OMG-Seg codebase
Other
905 stars 44 forks source link

Where is label of Object-365-instance segmentation #6

Closed ArsenLuca closed 3 months ago

ArsenLuca commented 3 months ago

as the title states

lxtGH commented 3 months ago

We use SAM-Huge to provide the mask label. Currently, we find the pre-training of such dataset only bring a little improvement (within 1% PQ on COCO dataset, less than 0.5% PQ on Cityscapes). You can also directly use CLIP-pretrained backbone for training OMG-Seg.