Question about reproducing the aliasing ratio analysis

[lqh305420-ops]

Hi, thank you for your excellent work.

I am trying to reproduce the aliasing ratio analysis shown in Fig. 1 and Fig. 5 of the paper. I understand that the aliasing ratio is defined as the ratio of spectral power above the Nyquist frequency to the total spectral power of a feature map. However, I am still unclear about some implementation details.

Could you please clarify how the aliasing ratio is computed in your experiments?

Specifically:

1. Which feature map is used for the analysis?
   Is it the feature before the first downsampling layer, or the feature before each 2× downsampling layer, such as the features between different backbone stages?

2. For multi-channel feature maps with shape [B, C, H, W], how are the channel dimensions handled?
   Do you compute the aliasing ratio for each channel and then average them, or sum the spectral power over all channels directly?

3. Is the aliasing ratio computed per image sample and then correlated with the per-image segmentation accuracy, as shown in the figures?

4. For models with multiple downsampling stages, do you report the aliasing ratio from a specific stage, or average the ratios across multiple stages?

5. How exactly is the “segmentation accuracy” in Fig. 1 and Fig. 5 computed?
   Is it pixel accuracy, per-image mIoU, or another metric?

It would be very helpful if you could provide the evaluation script or more details about this analysis.

Thank you!