Hi, I'm reading your paper and I'm a bit confused about one point.
In Section 2.1 of the paper, it is stated that there's a fixed proportion $r$ for each P-frame, i.e. $|Ω_{i,n,τ}|=⌊rP_0⌋$, to select salient patches, which suggests a local Top-K selection strategy. However, global Top-K is employed actually, as it is indicated in Section 8: "Top-k selection" "Patches are ranked globally by their saliency scores...." (also the code implementation). I also read #116 . It could be misleading for readers trying to reproduce your experiments.
Is it possible that the statement in Section 2.1 is inaccurate or outdated relative to the final implementation?
Hi, I'm reading your paper and I'm a bit confused about one point.
In Section 2.1 of the paper, it is stated that there's a fixed proportion$r$ for each P-frame, i.e. $|Ω_{i,n,τ}|=⌊rP_0⌋$ , to select salient patches, which suggests a local Top-K selection strategy. However, global Top-K is employed actually, as it is indicated in Section 8: "Top-k selection" "Patches are ranked globally by their saliency scores...." (also the code implementation). I also read #116 . It could be misleading for readers trying to reproduce your experiments.
Is it possible that the statement in Section 2.1 is inaccurate or outdated relative to the final implementation?