search

ZJULearning / resa

Implementation of our paper 'RESA: Recurrent Feature-Shift Aggregator for Lane Detection' in AAAI2021.
Apache License 2.0
175 stars 35 forks source link

Unable to find the feature aggregation with different strides aspect as mentioned in the paper #38

Closed sakethbachu closed 2 years ago

sakethbachu commented 2 years ago

Hello

I am unable to understand in the code how feature aggregation with different stride as mentioned in the paper (also in the picture present in the readme of this repo) is implemented. Any insight will be very helpful.

Thanks in advance!

Turoad commented 2 years ago

The implementation is in: https://github.com/ZJULearning/resa/blob/main/models/resa.py#L61-L71. You may focus on this code, it first gets the shifted feature with conv(x[..., idx, :] then adds to the original feature.

x.add_(self.alpha * F.relu(conv(x[..., idx, :])))

Please let me know if you have further questions.

Turoad commented 2 years ago

We will close this issue first. Feel free to reopen it if you have any questions.

2696120622 commented 1 year ago

The implementation is in: https://github.com/ZJULearning/resa/blob/main/models/resa.py#L61-L71. You may focus on this code, it first gets the shifted feature with conv(x[..., idx, :] then adds to the original feature.

x.add_(self.alpha * F.relu(conv(x[..., idx, :])))

Please let me know if you have further questions.

@Turoad What this self.alpha does? Is there some suggestions for setting its value?