Some questions about FLOPs about maskformer HOT 4 CLOSED

FLOPs comparison does not make any sense without specifying input sizes. We recompute all FLOPs with the same input size for a fair comparison. The FLOPs reported in the Swin Transformer paper use a different input size.

from maskformer.

In the appendix of Swin-Transformer: "Swin-T, Swin-S are trained on the standard-setting as the previous approaches with an input of 512×512. Swin-B and Swin-L with z indicate that these two models are pre-trained on ImageNet-22K, and trained with the input of 640×640".
In your paper, the crop size is also 512x512. What's the difference in the input size between Swin and Maskformer.

from maskformer.

Yes, we train our model using the same crop size as Swin Transformer. But FLOPs calculation is separate and does not need to use the same crop size as training. We confirmed with authors of Swin Transformer and they used different input size to calculate FLOPs. Thus, we re-calculate FLOPs for a fair comparison.

from maskformer.

from maskformer.