search

SHI-Labs / OneFormer

OneFormer: One Transformer to Rule Universal Image Segmentation, arxiv 2022 / CVPR 2023
https://praeclarumjj3.github.io/oneformer
MIT License
1.41k stars 128 forks source link

real-time or video segmentation #47

Closed wangchao0720 closed 1 year ago

wangchao0720 commented 1 year ago

**hello, can this structure be used for real-time segmentation or video segmentation**

praeclarumjj3 commented 1 year ago

Hi @wangchao0720, yes, you can use a lightweight backbone for real-time segmentation. Note that a Swin-L OneFormer outputs result at ~15 FPS on a single RTX A6000 GPU. So, a lighter backbone like ResNet-50 will give an even faster inference speed. You can perform video segmentation in a frame-by-frame manner.