search

berniwal / swin-transformer-pytorch

Implementation of the Swin Transformer in PyTorch.
https://arxiv.org/pdf/2103.14030.pdf
MIT License
794 stars 129 forks source link

Can this be used for target detection? I didn't make it #4

Closed kuozhijishou closed 3 years ago

kuozhijishou commented 3 years ago

Can this be used for target detection? I didn't make it

berniwal commented 3 years ago

Yes, the authors also used it for object detection by replacing the backbone of the respective object detector with the Swin Transformer. Depending on the object detector I assume you need to either extract the feature map after the last stage (before the global average pooling and linear layer) or for an FPN based approach extract the feature maps from each stage as they do for the ResNet. But in general just changing the backbones and leaving everything else the same should work.

Selection_376