Can torchvision.ops.RoIAlign be used instead of RoIAlign（CUDA extensions）

xlliu7 / E2E-TAD

[CVPR 2022] An Empirical Study of End-to-end Temporal Action Detection

Apache License 2.0

81 stars 11 forks source link

Can torchvision.ops.RoIAlign be used instead of RoIAlign（CUDA extensions） #5

Open alexandsunny opened 1 year ago

alexandsunny commented 1 year ago

I had some problems converting my model to onnx Can torchvision.ops.RoIAlign be used instead of RoIAlign（CUDA extensions）

xlliu7 commented 1 year ago

I guess it is okay after some modification. Note that torchvision.ops.RoIAlign only works for 2D images (4D feature maps of shape NxCxHxW). So you should resize the input feature sequence of the RoIAlign module from NxCxT to NxCx1xT and also modify the RoI locations to fit the new shape.